The Daily Bot

No. 026 · Friday, Nov 6, 2026

Note: The Daily Bot website is temporarily offline after hitting its hosting transfer limit. Until it is back, this email carries every story in full. Source links below still go to the original posts, docs and changelogs.

Today's one job is to ask a Cursor cloud agent to build an app and get a live share link: it can set up hosting, a database and Cursor sign-in, then hand you a URL. Also inside: Claude Personal Finance beta, official Grok Bot phone calls, Images 2.6 GIFs, Codex CLI 0.167, GPT-6 Cyber for select Daybreak Red orgs, Claude Enterprise hooks fine print, scientist usage research, a Hark travel-catch report, two morning workflows, and the Friday wrap on a Flare-draft, Sunburst-final GIF brand kit.

In this issue: Cursor live apps · What's new (6) · How people use it (2) · Workflows (2) · Opinion · Week in review · Digest · Known issues · Coming up · Daily Bot Check

Today's lead

Cursor cloud agents can now build, host and publish live apps (confirmed)

Ask a Cursor cloud agent for an app and it sets up the database, sign-in and hosting, checks it for leaked secrets, and gives you a link to share.

What to do:

1. Update Cursor and start a cloud agent, not a local one.
2. Say plainly that you want it to build an app, and describe who uses it and what data it shows.
3. Give it any secrets it asks for through the prompt Cursor shows. The agent never sees the values.
4. Preview the app in the tray, then choose Publish and pick its URL.
5. If a security check fails, tell the agent to fix what it lists, then publish again.
6. Share it with specific people, a group or your team. Public links work only on individual plans unless an admin turns them on.

Prompt (ours):

Build an app for my team that lists our open support tickets from the data warehouse in a table I can filter by owner and status, with a count of tickets opened this week at the top. Keep it team-only behind Cursor sign-in. Ask me for any secrets you need, and don't publish until I've previewed it. (ours, untested)

You can now ask a Cursor cloud agent to build a working web app and hand you a live link. Cursor sets up the hosting, a Postgres database if the app needs one, and sign-in, then runs security checks before anything goes live.

What changed

Cursor's changelog is dated Nov 5: "Now, you can build live apps directly in Cursor and share them with a link." It said Cursor will "build the app, set up a database, configure auth, and deploy it." The page was not on the changelog at a 2:20 PM ET scan (latest then was Oct 20), and it was live when we opened it at 4:20 PM ET, so it appeared in that window. The first third-party mention we found was @cursorlog at 3:55 PM ET: "Cursor Live Apps has been released." We did not find an @cursor_ai post for it in this gate.

Apps can come from new or existing repos on Cursor Origin or GitHub, and "Support for other git providers like GitLab is coming soon." Auth for private and team apps sits "behind Cursor sign-in," per the Apps docs. Each app runs in a hosting account Cursor creates for your team, "isolated from every other team's and user's." When an app needs a key, "The agent never sees the values."

Why it matters

Internal tools usually stall at "it works on my laptop." Cursor's own list of examples is mostly non-engineering: event sign-ups, sales pipeline dashboards, onboarding guides. A sales or ops lead can now get a team-only tool without filing a ticket for hosting and login. Grok Bot can also open a private app once you provision a Cursor API key, which is why this lands on our beat as well as Cursor's.

How to use it

1. Start a cloud agent. The docs say "this feature is not supported locally."
2. Say plainly that you want an app. Cursor's sample: "Deploy this repo as an app."
3. Preview it in the tray, choose Publish, and pick a URL.
4. Fix anything the checks flag, then publish again.

Our starter prompt (ours, untested):

Build an app for my team that lists our open support tickets from the data warehouse in a table I can filter by owner and status, with a count of tickets opened this week at the top. Keep it team-only behind Cursor sign-in. Ask me for any secrets you need, and don't publish until I've previewed it.

Good to know

• Availability: "Apps is available on all paid plans with access to cloud agents."
• Four checks always run before publish: they block secrets in files and credential files, and they leave out .env files and source maps. An AI review runs on top. Team admins can add review rules and can approve publishes the review blocked, except for malicious code, hardcoded secrets, or exposed credentials.
• Public sharing is "disabled by default for Teams and Enterprise." Admins can see every app and secret and unpublish any app. They can also turn Apps on or off, and allow or disallow API keys opening apps.
• Private apps open for a script or Grok Bot with a Cursor API key in the x-cursor-app-token header. The docs say "The key only works as a header." "There's no query parameter and no cookie." Our read: put the key only in your tool's request headers, not in a share URL or in client-side app code.
• Personal API keys open the apps their owner can open; team API keys open only the team apps they've been given access to.
• The docs say hosting is provided "as is" and "as available." No pricing or usage limits are listed on the changelog or Apps docs as of this check.
• We didn't try this hands-on.

What to watch

GitLab support, any usage or hosting limits, and whether Cursor posts the launch on X.

Last checked: Nov 5, 2026, 4:23 to 4:25 PM ET. What changed: first version; editor gate dated for Fri assemble, expanded sourced depth from Apps docs (SSO, four checks, admin controls, API key header rules), set Last checked from own page opens (writer ~400 words expanded for lead depth).

Sources

• Cursor changelog: Build Live Apps in Cursor
• Cursor docs: Apps
• Cursor docs: Cloud agent setup
• Cursor Changelog (@cursorlog) on X

What's new

No new Grok Bot version beyond v0.78.1 on the changelog as of this morning. The stories below are today's standalones from the Nov 6 issue.

Claude Personal Finance beta: connect banks, ask in plain words (Anthropic · confirmed)

US Pro, Max, and Ultra plans get a beta that links banks through Plaid so Claude can answer money questions from your balances and transactions, read-only.

What to do:

1. Our read: open Claude on the web, Desktop, or mobile, then open Personal Finance in the side menu (hover More if hidden).
2. Our read: select Get started, choose or skip goals, Continue, then Link accounts with Plaid and finish bank sign-in within 10 minutes.
3. Our read: wait for accounts to show, then ask on the Personal Finance page or in a chat with Plaid allowed.
4. Our read: before trusting a figure, ask which transactions sit behind it, and check the number in your bank app.

Prompt (ours):

Using only my linked accounts, total last month's spending by category, flag duplicates or unexpected fees, and say what is left before payday. Cite the transactions behind each number. Do not move money or give investment advice. (ours, untested)

If you are on Claude Pro, Max, or Ultra in the US, Anthropic says you can open Personal Finance, connect bank and card accounts through Plaid, and ask about your money in your own words. Access is read-only. Claude cannot move, manage, or invest your money.

What changed

On Nov 5, Anthropic published Introducing Personal Finance in Claude, listed it on the product announcements index, and @claudeai posted at 1:00 PM ET: "Introducing Personal Finance in Claude, in beta in the US." The article says: "Personal Finance starts rolling out in beta this week, on the Pro, Max, and Ultra plans in the US." The Help Center covers web, Desktop, iOS, and Android, excludes Free, Team, and Enterprise, and warns: "We're rolling it out gradually, so you might not see it yet."

You sign in at your bank through Plaid. The article says Claude does not ask for or see your bank password. For connected accounts, Anthropic says Claude sees the bank name, each account's name, type, and last four digits, balances and credit limits, and up to two years of transactions. The Help Center words the history as "Up to 24 months of transactions, including the amount, date, merchant, category, and your bank's description."

Why it matters

For a US Pro, Max, or Ultra subscriber, this is one place to ask across the accounts you linked, instead of opening every banking app yourself. Anthropic says Personal Finance has no ads or sponsored recommendations, and that it does not offer credit cards, loans, or savings accounts of its own. Hub chats are carved out of training: "When you chat with Claude in the Personal Finance hub, those conversations aren't used to train our models." Outside the hub, your privacy settings apply. This confirms the Oct 19 unreleased Plaid sighting.

How to use it

Our read from the Help Center setup steps (untested: US paid beta only; no Plaid-linked US account on the box):

1. Open Claude on the web, Desktop, or mobile, then open Personal Finance in the side menu (hover More if hidden).
2. Select Get started, choose or skip goals, Continue, then Link accounts with Plaid. Finish bank sign-in within 10 minutes and return to Claude right away.
3. Wait for accounts to show, then ask on the Personal Finance page or in a chat with Plaid allowed.
4. Before trusting a figure, ask which transactions sit behind it, and check your bank app.

Copyable prompt (ours, untested):

Using only my linked accounts, total last month's spending by category, flag duplicates or unexpected fees, and say what is left before payday. Cite the transactions behind each number. Do not move money or give investment advice.

Gotchas

Sourced from the Help Center unless noted. US banks only; not on Free, Team, or Enterprise; outside the US or on a foreign VPN the sidebar stays hidden. Cap is 16 banks. "Balances can be a day or two old." "If your bank shows it differently, it may be counted twice in your totals." Claude reads balances and credit limits, not holdings, trades, rates, or terms. Plaid usually shares only the last few digits, but "If your bank writes a full number into an account's name or a transaction's description, Claude may see it there." Linked accounts do not work in scheduled tasks, Cowork tasks, or Claude Code, and may not work in a project; not available in Claude Duet. Shared chat links and published artifacts can expose account figures. Unlinking your last bank resets the dashboard and setup answers, and that cannot be undone. Claude can make mistakes; the article says Personal Finance is not designed to give investment advice, so check important figures with your bank.

What to watch

Anthropic says it is exploring how Claude can work alongside licensed professionals and regulated firms. Our read: watch for a Help Center or product post that names a bank, wealth manager, or advisor partner, or that documents a professional handoff inside Claude.

Last checked: Nov 5, 2026, 3:53 to 3:56 PM ET. What changed: first version; editor gate set Last checked from own Anthropic and Help Center opens (writer end matched gate start), dated entry for Fri assemble, fixed balances quote punctuation.

Sources

• Introducing Personal Finance in Claude (Claude Resources, Nov 5, 2026)
• Use Personal Finance in Claude (Claude Help Center)
• Product announcements index
• @claudeai on X: Introducing Personal Finance in Claude (Nov 5, 2026)
• TechCrunch: Claude can now analyze your spending with its new Personal Finance tool
• Unreleased Claude screens show Personal Finance with Plaid bank linking (The Daily Bot, Oct 19)

It's official: Grok Bot can make phone calls (Grok Bot · confirmed)

Grok Bot's own account announced outside calls on Nov 4, so this is no longer just user reports, though the changelog and plan details still haven't caught up.

What to do:

1. Tell your Bot who to call and the one thing the call should get done.
2. Say what it may and may not agree to, and what it may share.
3. Check the number on the approval card, then choose Call.
4. Watch the live transcript. Use Take over or End call if it goes off track.
5. Read the result and replay the recording before you act on it.

Prompt (ours):

Call my internet provider and ask what it would cost to drop to the cheapest plan. Don't cancel, accept an offer, or agree to anything on the call, and don't share my account PIN or card details. Show me the number before you dial. After the call, send me the price they quoted in their words. (ours, untested)

Your Bot can place a phone call for you, and now Grok Bot says so itself. At 1:13 PM ET on Wednesday the official @bot account posted: "Grok Bot can now make phone calls."

What changed

The post carries a 42-second promo video. In it the Bot says "Your internet contract renews in 3 days" and asks "Want me to call and cancel?" The user replies "And don't let them talk me into a discount."

Before dialing, an approval card reads "Call your internet provider?" with the number and Cancel and Call. During the call a live transcript shows each side, with Take over and End call. When the agent offers "10% off for a year if you stay," the Bot declines. Afterward a recording bar appears and the Bot reports "Done. Your plan is canceled."

Elon Musk followed up at 2:59 AM ET Thursday: "Grok @Bot can call your dentist on the phone!" He quoted a user whose Bot was "calling a dentist for me to see if they take my insurance."

Why it matters

Hold music, cancellations and "call us for details" can go to your Bot. The approval card and Take over button keep you in control.

How to use it

Use the steps in the card, or start from this prompt (ours, untested):

Call my internet provider and ask what it would cost to drop to the cheapest plan. Don't cancel, accept an offer, or agree to anything on the call, and don't share my account PIN or card details. Show me the number before you dial. After the call, send me the price they quoted in their words.

Good to know

• The changelog still tops out at v0.78.1 (Oct 27) with no calls entry. The video names no plans, countries or usage cost.
• The video is a promo, and the number on its card is a 555 number. Decide yourself whether a Bot should cancel or accept anything for you.
• We reviewed the promo frame by frame. We have not placed a call ourselves.

What to watch

A changelog or help page entry with availability, limits and usage per call.

Last checked: Nov 5, 2026, 10:16 to 10:22 AM ET (@bot/Musk/Arthur posts via X API; video OCR; changelog at v0.78.1). What changed: first published for Fri Nov 6; upgrades earlier reported calls coverage to confirmed.

Sources

• Grok Bot on X: calls announcement
• Elon Musk on X: dentist call
• Arthur MacWaters on X
• Grok Bot changelog
• The Daily Bot: calling a business for missing details

OpenAI Images 2.6 adds GIFs and faster API models (OpenAI · confirmed)

ChatGPT, Codex and the API get native GIFs, reusable reference images, 50% faster generation, and GPT-Image-2.6 Flare and Sunburst.

What to do:

1. In ChatGPT or Codex, create with Images 2.6 (available on all subscription tiers on desktop, mobile and web per OpenAI).
2. For API GIFs, call image generation with model gpt-image-2.6-flare or gpt-image-2.6-sunburst and set output_format to gif, per OpenAI's image generation guide.
3. For API edits, send a transparent markup layer on supported models and set complexity to low for quicker edits when the scene is simple, per the same guide.

You can make animated GIFs in ChatGPT now, and developers get two new image models. Per OpenAI's announcement, "We're introducing Images 2.6 in ChatGPT, Codex, and the API."

What changed

OpenAI said Images 2.6 brings "GIFs and reference images" plus "two new models" for the API "with 50% faster image generation." In the same thread, OpenAI added: "Keep reference images of yourself, pets, and characters ready to reuse in new creations" and "no need to upload them again." The announcement adds GPT-Image-2.6 Flare and Sunburst, API image markup, SVG export, and a complexity parameter. The Nov 4 API changelog lists the same capabilities for both models.

Why it matters

Our read: ChatGPT finally gets a native GIF path, and the API split is clear. The announcement says "Images 2.6 generates images 50% faster than before," and "When quality is set to low, Flare can generate images in as little as six seconds." Sunburst is the higher-detail, edit-consistency option. Per OpenAI's Flare and Sunburst model pages, pricing matches: text input $5/1M tokens, cached text $1.25/1M; image input $8/1M, cached image $2/1M; image output $30/1M.

How to use it

1. In ChatGPT or Codex, use Images 2.6. Per the announcement, "Images 2.6 is available globally today in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web."
2. For API GIFs, per OpenAI's image generation guide, set model to gpt-image-2.6-flare or gpt-image-2.6-sunburst on POST /v1/images/generations and set output_format to gif.
3. For pinpoint edits, send a transparent markup layer; set complexity to low, medium, or high (default high if omitted). OpenAI recommends low complexity "for most editing tasks."

Ours, untested, for a ChatGPT GIF:

Make a looping GIF of a desk lamp clicking on and off. Keep the camera fixed, 6 frames, no text.

Good to know

• Per the image generation guide, GIF and SVG output are generation-only; the edit endpoint does not support them. Do not combine markup with mask or prompt_rewriter. Complexity has no separate fee. "50% faster" baseline unnamed. No hands-on (Luke has no ChatGPT).

What to watch

Compare Flare vs Sunburst on edit loops, and check whether ChatGPT surfaces saved reference images the way the X thread describes.

Last checked: Nov 5, 2026, 12:02 to 12:15 PM ET. What changed: first filed Nov 5 after the Nov 4 evening release was missed during the OpenAI Desk pause.

Sources

• OpenAI: Introducing Images 2.6 (Nov 4, 2026)
• OpenAI API: Image generation guide
• OpenAI API: GPT-Image-2.6 Flare
• OpenAI API: GPT-Image-2.6 Sunburst
• OpenAI API changelog (Nov 4, 2026)
• OpenAI on X: ChatGPT Images 2.6

Codex CLI 0.167 opens the agent command center without a daemon (OpenAI · confirmed)

Update to open the agent command center with unread marks, switch backends while keeping drafts, preview local images, and set up Vertex AI or a system proxy.

What to do:

1. Update with npm install -g @openai/[email protected], or the install script / Homebrew upgrade path in OpenAI's Codex CLI docs.
2. Open the agent command center with codex agents --no-daemon; unread results show in the command center and session picker, per the release notes.
3. Switch backends from the terminal session picker; drafts are kept when you switch, per the release notes and PR 58457.
4. To try Vertex AI setup, enable features.gcp_setup_wizard (off by default) and use Sign in another way, per the release notes and PR 57857.
5. For macOS or Windows system proxy settings, open /experimental and enable respect_system_proxy, then restart Codex, per PR 58157.

If you juggle several Codex backends or embedded sessions, update to 0.167.0. The release notes (published Nov 5, 2026, 8:40 PM ET; listed under Nov 6 on the ChatGPT & Codex changelog) open with: "Open the agent command center with codex agents --no-daemon. The command center and session picker now mark unread results".

What changed

Before 0.167.0, PR 58238 says command-center navigation "required a shared app server, leaving embedded sessions without the same way to switch between chats." This release ships that path plus unread marks. You can also "Switch between connected backends from the terminal session picker and keep your drafts when switching", per the release notes. PR 58457 says selecting another backend's chat used to be rejected. Local image thumbnails appear while composing "on supported terminals," with clickable viewed-image paths. Sign-in adds API key and cloud options under "Sign in another way", including a Vertex AI wizard. Opt-in system proxy support for macOS or Windows lands under /experimental.

Why it matters

Our read: multi-backend and embedded users get one command center with unread marks and draft-safe hops, so a half-written prompt survives a session jump.

How to use it

1. Update with npm install -g @openai/[email protected], or follow OpenAI's Codex CLI docs install path for your platform.
2. Run codex agents --no-daemon (release notes).
3. Switch backends from the session picker; drafts stay (release notes; PR 58457).
4. Optional: enable features.gcp_setup_wizard, then use "Sign in another way" for Vertex AI (PR 57857).
5. Optional: in /experimental, enable system proxy (respect_system_proxy), then restart (PR 58157).

Ours, untested:

From the agent command center, open the unread session that finished overnight. Summarize what changed, list open approvals, and draft the next prompt without submitting it.

Good to know

• gcp_setup_wizard is "disabled by default," per PR 57857.
• System proxy stays off until you opt in; PR 58157 says to restart after changing it.
• Image previews need a "supported" terminal, per the release notes.
• PR 58457: backend switches are blocked "during an active turn", and a failed destination prep keeps the current conversation.

What to watch

Watch OpenAI's Codex CLI docs pick up codex agents --no-daemon, unread marks, and the Vertex / proxy flags. npm resolves @openai/codex to 0.167.0 as of this check, and the alpha tag already points at 0.168.0-alpha.7, the next build to watch (npm).

Last checked: Nov 6, 2026, 6:02 to 6:06 AM ET. What changed: first version of this entry, filed after 0.167.0 moved from alpha builds to the stable release on GitHub and npm.

Sources

• Codex CLI 0.167.0 release notes (GitHub)
• ChatGPT & Codex changelog (Nov 6, 2026)
• Codex pull request 58238: Enable the agent command center for embedded sessions
• Codex pull request 58457: Enable cross-backend session switching in the TUI
• Codex pull request 57857: Add Google Cloud Vertex AI setup to TUI onboarding
• Codex pull request 58157: Expose system proxy support as an experimental setting
• OpenAI docs: Codex CLI (install and upgrade)
• npm: @openai/codex package

OpenAI GPT-6 Cyber ships for select Daybreak Red orgs (OpenAI · confirmed)

Approved Daybreak Red organizations with separate model approval can call gpt-6-cyber in the API and Codex; Red alone is not enough.

What to do:

1. Confirm Daybreak Red and separate GPT-6 Cyber model approval with your OpenAI account team, per OpenAI's Daybreak guide.
2. Have an organization owner enable Daybreak Red on a non-default API project, set an IP allowlist, and configure Private Safety Processing if the project uses Zero Data Retention, per the same guide.
3. Call the Responses API with model gpt-6-cyber and access_programs.cyber set to daybreak_red, using a key from that project, per OpenAI's Daybreak guide.

If your org already has Daybreak Red and separate GPT-6 Cyber approval, you can call gpt-6-cyber in the API and Codex today. Per OpenAI's announcement, "GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers."

What changed

OpenAI said, "Introducing GPT-6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment," available to "select Daybreak Red customers," expanding over time. The announcement rates it "Critical in cybersecurity under our Preparedness Framework, up from High for GPT-5.6 Cyber," with ExploitGym 48.5% vs 32.1% for GPT-5.6 Cyber. The Nov 5 API changelog says to use it "to investigate vulnerabilities, reproduce them in an authorized test environment, and verify fixes." OpenAI is also "making GPT-5.6 Cyber available to all Daybreak Red customers."

Why it matters

Our read: gated specialist, not ChatGPT default. The Help Center says "GPT-6 Cyber is initially available only to Daybreak Red organizations approved for that model," while "All Daybreak Red organizations are eligible for GPT-5.6 Cyber." Per the model page, short-context pricing is $20 input, $2 cached input, $25 cache writes, and $100 output per 1M tokens.

How to use it

1. Confirm Daybreak Red plus separate GPT-6 Cyber approval with your account team (Daybreak guide: "Confirm with your account team that your organization has Daybreak Red access and approval for GPT-6-Cyber.").
2. Enable Daybreak Red on a non-default project, set an IP allowlist, and configure Private Safety Processing if using Zero Data Retention, per the same guide.
3. Call Responses with model gpt-6-cyber and access_programs.cyber daybreak_red from that project's key. OpenAI's official sample input: "Suggest a regression test for a path traversal fix."

Ours, untested:

Summarize high-severity findings in this authorized scan report. For each, give title, path, and a regression test idea. Stay inside scoped paths.

Good to know

• Per the Help Center: "Turning on Daybreak Red in your workspace or API project doesn't grant that approval."
• Per the Help Center: "Selecting Daybreak Red doesn't change your model or grant access to another one." Set model to gpt-6-cyber.

What to watch

Watch OpenAI expand GPT-6 Cyber past the initial select set. Nov 12: "Cyber at the Frontier" livestream at 10am PT (X).

Last checked: Nov 5, 2026, 4:02 to 4:05 PM ET. What changed: first version; editor gate set Last checked from own OpenAI page opens, dated for Fri assemble, confirmed Critical is cybersecurity-only under Preparedness Framework, cover re-exported greyscale, difficulty Power user (Advanced invalid).

Sources

• OpenAI: GPT-6 Cyber (Nov 5, 2026)
• OpenAI API changelog (Nov 5, 2026)
• OpenAI: Use Daybreak in the Responses API
• OpenAI API: GPT-6-Cyber model page
• OpenAI Help Center: Daybreak overview
• OpenAI on X: Introducing GPT-6 Cyber

Claude Enterprise hooks: what they block, and what slips past (Anthropic · confirmed)

Anthropic added a beta security center Nov 5. Hooks, out of beta since Oct 20, check prompts before Claude runs, but skip images and default to fail open.

What to do:

1. If you use Claude at work: ask your admin whether inference hooks are on, whether failure handling is set to block or allow, and whether your role is excluded.
2. Do not treat a screenshot or image of sensitive data as covered: hooks never receive raw image bytes.
3. Admins: start in shadow mode, then choose Block the request deliberately; the first-save default is Allow the request.
4. Admins: pair hooks (before the model runs) with the security center report and Compliance API (after the fact).

Prompt (ours):

Using only the security center gaps listed below (current value, recommended value, setting link), draft a one-page Owner review: which gaps to close this week, which to accept with a written risk note, and which need another team. Do not change any settings. (ours, untested)

If your company runs Claude Enterprise, what you type into Claude at work may now be checked by your company's security tools before the model ever sees it. On Nov 5 Anthropic added a beta security center page and bundled it with inference hooks, which were already out of beta. The useful part for readers is the fine print: hooks skip images, can be set to let requests through when the checker is down, and do not cover every Claude surface.

What changed, and when

• Aug 5, 2026: the Platform release notes list inference hooks in beta for Claude Enterprise organizations.
• Oct 20, 2026: the same release notes say inference hooks "are out of beta for Claude Enterprise organizations."
• Nov 5, 2026: Anthropic's announcement describes two kinds of hooks, prompt hooks and pre-tool-call hooks, as "both now generally available" and packages them with new beta features. The Claude Apps release notes list the same package, "several in beta." We found no dated primary saying either hook type first became generally available on Nov 5, so we treat hooks as a restated GA feature, not new news.

What is new on Nov 5 is the security center (beta). The announcement says security teams can now see chat, Cowork, and Claude Code settings "on one page, instead of setting by setting", with the current value, Anthropic's recommended value where there is one, and a link to change it. Other Nov 5 betas: a parent-organization console at admin.claude.com, Admin API sign-out, a self-serve IP allowlist (before, the announcement says, admins went "through Anthropic"), and Claude Desktop controls that can block personal Claude accounts on managed computers.

Why it matters

The three tools act at different moments, and only one of them can stop anything.

• Hooks act before the model runs. Per the technical docs, Anthropic sends the transcript to the company's security server and waits for allow or deny; "a denied request never reaches the model." With Validate tool calls on, tool calls are checked before they run.
• The security center is a report. It shows how settings compare with Anthropic's recommendations. The announcement says it is "not a certification or a guarantee of compliance with any law or standard." It blocks nothing.
• EDR in Cowork watches, it does not stop. The announcement says sensors in the local Cowork virtual machine "report activity to your security team and don't block it" and cover "the virtual machine on the user's computer, not Cowork sessions in the cloud."

So for an employee, a paste of customer data into a governed chat can be refused on the spot, but a screenshot of the same data is not inspected, and a configured local Cowork EDR sensor reports activity without blocking it.

What hooks do not catch

The Help Center overview summarizes coverage as Claude, Claude Code, Cowork, and other Claude Enterprise products. The technical docs are narrower; they apply to "governed requests":

• Images: raw file and image bytes are never sent, "so image-only content (for example, a screenshot of a document) is not inspected." Attachments arrive only as metadata and extracted text.
• Out of scope: Claude Platform API organizations, and Claude on Amazon Bedrock or Google Cloud.
• Hosted features: Claude Security scans, Code Review on GitHub pull requests, and smart reports make their own model calls that are not sent to your endpoint.
• Some tool calls and voice: some claude.ai built-in tool calls (for example, searching connectors or reading past chats) may be left out of tool-call checks, and in voice mode "a small share of tool calls isn't" sent, so the docs say not to rely on hooks as the only voice control.
• Fail-open vs fail-closed: if the security server is down, slow, or erroring, the configuration docs let admins choose Block the request (fail closed) or Allow the request (fail open, no inspection). On first save the defaults are Allow the request and a 5,000 ms timeout.
• Partial rollout: requests outside a sampled rollout percentage "proceed without inspection, even when failure handling is set to Block the request", and admins can exclude custom roles. Shadow mode blocks nothing.
• No redaction: verdicts are only allow or deny.

Other limits in the package

• Centralized management: per the announcement, "Setup can't be undone, so identify your parent organization admins before you start." A parent role does not grant access inside an organization by default, but parent Compliance API keys can read data across organizations.
• IP allowlist: the Help Center says artifacts shared outside your organization aren't limited by the allowlist.
• Admin API sign-out does not end API keys, cloud sessions already running, or scheduled routines, per the announcement. Our opinion column "Opinion: Claude sign-out needs a receipt for what stays active" covers that gap.
• EDR in Cowork is in beta on macOS and Windows for Claude Team and Enterprise, and for the Claude desktop app on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, deployed through MDM for supported vendors, per the announcement.

How to use it

These steps are ours, from the docs; we have not configured hooks or opened the security center.

1. If you use Claude at work, ask your admin three things: are inference hooks on, is failure handling Block or Allow, and is your role excluded.
2. Treat screenshots and images of sensitive documents as uninspected by hooks.
3. Admins: run shadow mode first, then pick Block the request on purpose instead of keeping the Allow default.
4. Admins: use hooks to stop requests before the model runs, and the security center report or Compliance API to review afterward.

Copyable review prompt (ours, untested):

Using only the security center gaps listed below (current value, recommended
value, setting link), draft a one-page Owner review: which gaps to close this
week, which to accept with a written risk note, and which need another team.
Do not change any settings.

What to watch

• A Platform or Claude Apps release-note entry moving the security center or centralized management out of beta.
• Changes to the "Current limitations" list in the technical docs, especially image inspection or redaction.
• The supported EDR vendor list in the Help Center, and whether cloud Cowork sessions get EDR coverage.

Last checked: 2026-11-06, 6:10 AM to 6:13:22 AM ET What changed: Nov 6 refile after Editor Job KP. Added the Aug 5 / Oct 20 hooks chronology, the technical scope limits (governed requests, image-only blind spot, fail-open default, excluded surfaces), a Why it matters section, and a redrawn cover; removed unmerged related links. Editor qualified the EDR example to a configured sensor and removed the unsupported imageCaption field; the site supplies the Illustration caption.

Sources

• More visibility and control over Claude Enterprise for security teams (Claude Resources, Nov 5, 2026)
• Inference hooks (Claude Platform docs)
• Configure Inference hooks (Claude Platform docs)
• Claude Platform release notes
• Inference hooks overview (Claude Help Center)
• Restrict access to Claude with IP allowlisting (Claude Help Center)
• Claude Apps release notes

How people are using it

Setups people say they run, plus one research read. Self-reported items are labeled; we have not run them. (The shared Data Bot playbook was covered in full in the Nov 5 email; skipped here.)

Anthropic: scientists use Claude most for data analysis and writing (Anthropic · confirmed)

In 207,049 consumer-plan research sessions, data analysis and writing led, while ideation and data gathering, which scientists rank most important, trailed.

What to do:

1. Our read: list one week of your research work and sort each item into the report's eight task categories (data analysis, writing and communication, data gathering, reading and knowledge acquisition, theoretical analysis, design and development, ideation and hypothesis generation, other).
2. Our read: start with the tasks the report says get the most use relative to time, such as visualizing data and applying statistical tests.
3. Our read: if your work is code-heavy, try those tasks in Claude Code, where the report found data analysis is 48.1% of research sessions.
4. Our read: if you work in biology or biomedicine and have a paid plan, compare Claude Science on a database-heavy task, where the report found more ideation and more autonomous use.

Prompt (ours):

Here is a list of what I did on my research project this week. Sort each item into one of these categories: data analysis, writing and communication, data gathering, reading and knowledge acquisition, theoretical analysis, design and development, ideation and hypothesis generation, other. For each one, say whether you could do it, help with it, or can't help (for example, physical lab work), and what you'd need from me. Then pick the two items where handing you the work would save me the most time. (ours, untested)

If you do research with Claude, Anthropic's new study of 207,049 sessions says scientists lean on it most for data analysis and writing, the kinds of tasks where Anthropic's own capability measure scores AI highest. You can sort your own week into the report's eight task categories and pick what to hand off first. Ideas and data gathering, which scientists rank most important, get far less use.

What changed

On Nov 5, Anthropic's Economics team published AI for scientific discovery: How researchers use Claude. It labels 207,049 research sessions on consumer plans from August 2 to September 30, 2026: 128,914 from Chat and Cowork (which the report calls "Claude") and 78,135 from Claude Code. A classifier tagged each session with a field and one of 7,262 research tasks. Anthropic says only those labels were kept. A survey of 528 scientists sits alongside the session data.

To judge whether use is high or low, the report compares each task category's share of sessions with the share of time researchers typically spend on it. Those time shares are estimates from SciNet, the task taxonomy the study is built on, not a measurement of the users in the sample.

What it found

• Data and writing lead. "Scientists primarily use Claude to analyze data and to write," the report says. Data analysis is 33.7% of Claude sessions and 48.1% of Claude Code sessions, against an estimated 27.6% of researcher time. Writing is 32.0% of Claude sessions, against 16.0% of time. Drafting manuscripts is the single most frequent task: 18.6% on Claude and 11.1% on Claude Code.
• Ideas and data gathering trail. Data gathering takes 28.6% of researcher time but 11.5% of Claude sessions. Ideation is 4.2% of Claude sessions and 2.1% on Claude Code, against 7.1% of time. Survey respondents rank ideation first and data gathering second for producing important results.
• Coding fields dominate Claude Code. Computer science is 4.3% of the world's researchers but 32.7% of Claude Code sessions.
• Use follows capability. Tasks in the top 20% of Anthropic's AI capability score get about twice the expected use on Claude and about four times on Claude Code. The report puts it this way: "In short, scientists appear to use AI where it works best."
• Claude Science looks different. On Claude Science, ideation is 8% of sessions, which the report calls roughly on par with the share of researcher time it takes. In biology and biomedicine, over 70% of its sessions run with high AI autonomy, against 23.7% on Claude and 64.1% on Claude Code.

Why it matters

Our read: for anyone doing research, this is Anthropic's own map of where Claude already carries real work (analysis, writing, code) and where it barely shows up (ideas, physical data collection). That tells you which tasks to try handing off first, and where a specialized tool may serve you better than the general chat.

Anthropic also models how much faster research could go. With current capabilities, widely adopted, it estimates AI could take on 27% to 52% of the researcher time a study needs (median 38%), and output could rise from about 1.35x in the arts to 1.96x in computer science and 2.02x in statistics. If AI handled all non-physical work, Anthropic says fields with little lab work could reach about 15 to 16 times their output, while bench sciences reach 3 to 4 times. About 40% of survey respondents expect AI to carry out full research projects by 2030, and 16% say it will never happen.

How to use it

These steps are our reading of the report, not Anthropic's advice, and we haven't run them.

1. List one week of your research work and sort each item into the report's eight task categories: data analysis, writing and communication, data gathering, reading and knowledge acquisition, theoretical analysis, design and development, ideation and hypothesis generation, other.
2. Start with the tasks the report says get the most use relative to time, such as visualizing data and applying statistical tests.
3. If your work is code-heavy, try those tasks in Claude Code, where the report found data analysis is 48.1% of research sessions.
4. If you work in biology or biomedicine and have a paid plan, compare Claude Science on a database-heavy task, where the report found more ideation and more autonomous use.

A prompt to start the sort (ours, untested):

Here is a list of what I did on my research project this week. Sort each item into one of these categories: data analysis, writing and communication, data gathering, reading and knowledge acquisition, theoretical analysis, design and development, ideation and hypothesis generation, other. For each one, say whether you could do it, help with it, or can't help (for example, physical lab work), and what you'd need from me. Then pick the two items where handing you the work would save me the most time.

Good to know

• Scenarios, not forecasts. Anthropic's words: "These are illustrative scenarios, not forecasts."
• The time baseline is an estimate. The researcher time shares come from SciNet. The report says they are "based on an LLM classifier, validated using both wet-lab research protocols that record researcher time spent per task and the frequency of tasks across thousands of academic papers."
• Cost drives the ceiling. The model assumes an hour of AI agent time costs $30 against $95 for a researcher hour. For fields without physical bottlenecks in its fully automated case, Anthropic says halving the cost of AI would double the rate of progress.
• Ideation may be undercounted. "Our data may also undercount ideation, since it tends to be woven into many other research tasks and is difficult to classify."
• Who's in the survey. Respondents applied to Anthropic's AI for Science and related grant programs. "Because they applied for AI research credits, they are likely more intensive AI users than the typical scientist."
• Claude Science users self-select. Anthropic notes adopters are likely those with more demanding use cases: "We therefore describe the differences across products as patterns of use."
• Consumer plans only. The report says its main sample is sessions on consumer plans, over about two months.
• Watch next. Our read: the sign that the ideation gap is closing is a rising ideation share on Claude Science or Claude in a follow-up report.

Last checked: Nov 5, 2026, 11:48 to 11:55 AM ET. What changed: first published Nov 6; every figure and quote rechecked against the Anthropic report on Nov 5.

Sources

• AI for scientific discovery: How researchers use Claude (Anthropic, Nov 5, 2026)
• Anthropic Research index
• Claude Science (beta) product page

A user says Hark caught hotel and flight mismatches (Hark · reported)

Dale Thomas says Hark flagged a Saturday hotel checkout before a Sunday flight, then reported three travel mismatches it found on the free plan.

What to do:

1. Connect the calendar and email Hark should read for trips, if you have not already.
2. Ask Hark to audit upcoming trips for hotel, flight, and car-rental mismatches.
3. Read the list of issues before you approve any booking or payment.
4. Confirm only the nights or changes you want; tell Hark not to spend until you say so.
5. Watch your usage if you are on the free plan, since proactive scans can burn the monthly cap.

Prompt (ours):

Look at my upcoming trips for the next three months. Check hotels, flights, and car rentals against each other and against my calendar. List every mismatch (checkout before a flight, rental ends before arrival, missing nights). Do not book, pay, or change anything yet. Put the list and your questions at the top of your reply, and wait for me to confirm before you add a hotel night or spend money.

What you can do now: ask Hark to audit your upcoming trips for hotel, flight, and car-rental mismatches, then confirm before it books or pays for an extra night. Dale Thomas says that is how Hark caught a Saturday hotel checkout before a Sunday flight while he was only booking a haircut.

What changed / the report

On Thursday morning, Dale Thomas (@dalethomas50) wrote at 7:48 AM ET: "Learned of @hark_labs through @brian_singer and Hark just saved my butt. While booking a haircut appointment for me, it alerted me that on one of my trips, I checked out of my hotel on Sat., but my flight isn't until Sun!!! Hark is adding the night now." He closed: "Hark is amazing." (user report)

About two and a half hours later, Thomas followed up at 10:23 AM ET with a longer note: "@hark_labs just crushed it again! I have a ton of travel in the next few months with the family. Between hotels, car rentals, and flights, sometimes it gets a little nuts." Then: "Hark found 3 separate issues where hotels, flights, and car rentals didn't align. Mind you, it found these proactively. And I am only on the free version ... but upgrading!" (user report)

Brett Adcock, Hark's founder, replied at 2:59 PM ET: "Love this" (founder post). Hark's company account wrote: "Death, taxes, and travel getting nutty. Hark's got you, Dale!" (company post)

We did not repeat the audit. Thomas's posts are a single user's report of what Hark did on his account.

Why it matters

A missed hotel night before a morning flight is a real cost: either you sleep somewhere else, or you pay for the night when you notice too late. Thomas says Hark spotted the Sat. checkout vs Sun. flight while he was mid-haircut booking, then reported three hotel, flight, and car-rental mismatches it found without him asking for a trip review. That is the proactive catch users hope for from an assistant that already sees calendar and email. Our Nov 3 Mac three-task test showed Hark reading a bus ticket from email while planning a trip; Thomas's thread is the travel-audit version in the wild.

How it maps to Hark's docs

Hark's Using Hark page says Action Buttons are "Generated by Hark" from "sources like what Hark sees in your email and has learned from your conversations," and that Handoff is "a computer in the cloud" that "can sign into websites for you and take actions (like buying things and filling out web forms) on your behalf." The same page lists "Manage your calendar" and "Book flights" among everyday jobs, and says memory improves the longer you use Hark. Pricing still says the free plan includes the full feature set, with Pro² at $20/mo and Pro³ at $100/mo for more usage. Thomas's "adding the night" line fits Handoff booking; confirm before you let that spend go through.

How to use it

1. Connect the calendar and email Hark should read for trips, if you have not already.
2. Ask Hark to audit upcoming trips for hotel, flight, and car-rental mismatches.
3. Read the list of issues before you approve any booking or payment.
4. Confirm only the nights or changes you want. Tell Hark not to spend until you say so.
5. Watch your usage if you are on the free plan. Thomas said he is upgrading after the free-plan catches.

A prompt to try (ours, untested):

Look at my upcoming trips for the next three months. Check hotels, flights, and car rentals against each other and against my calendar. List every mismatch (checkout before a flight, rental ends before arrival, missing nights). Do not book, pay, or change anything yet. Put the list and your questions at the top of your reply, and wait for me to confirm before you add a hotel night or spend money.

Good to know

• Confirm before Hark spends. Thomas wrote that "Hark is adding the night now." Treat hotel nights and car changes as paid actions. Approve the night, the rate, and the card before Handoff books.
• Free-plan limit risk. Thomas said he is "only on the free version ... but upgrading!" Free usage still hits caps; see our Nov 2 known-issue on Dec 1 reset reports if your meter looks wrong. This piece is not a re-lead of that story.
• Single-user report. We have Thomas's posts and the company replies. We have not verified the three mismatches against his bookings.
• Adcock is Hark's founder. His "Love this" reply is a founder post, not an independent review.

What to watch

Whether Hark publishes a travel-audit Specialty or Panel for hotel, flight, and car-rental alignment, and whether more free-plan users report the same proactive catches without burning the month.

Last checked: Nov 6, 2026, 6:14 to 6:16 AM ET. What changed: corrected weekday (Dale's posts are Thursday Nov 5, not Wednesday); tightened three-mismatch wording so it matches Dale's posts (no "two more" inference).

Sources

• Dale Thomas (@dalethomas50) on X: hotel checkout Sat vs flight Sun (user report)
• Dale Thomas on X: three hotel, flight, and car-rental mismatches (user report)
• Brett Adcock (@adcock_brett) on X: Love this (founder post)
• Hark (@hark_labs) on X: Death, taxes, and travel getting nutty (company post)
• Hark: Using Hark Pro (Action Buttons, Handoff, calendar, memory, usage)
• Hark: Pricing (free full features, Pro² $20, Pro³ $100)

Workflows and Setup of the day

A ChatGPT scheduled task drops the morning news into chat (OpenAI · reported)

One reader's every-morning ChatGPT task delivers the news, then forces a scroll hunt to find today's run.

You can turn ChatGPT into a morning news drop with a scheduled task, then open the same chat for the brief. althie krow (@AlthieMedia) wrote on Nov 3: "I use ChatGPT as my RSS reader. A scheduled task, every morning, my news."

What it replaces

Opening a feed reader or a stack of news sites first thing. The author's point is that the chat is doing the pull, not that the reading UX matches a feed.

How it works

1. A ChatGPT scheduled task runs every morning and writes the news into a chat.
2. The author opens that chat for the day's drop.
3. The gotcha is the delivery surface: "Every morning it drops me at the bottom of the chat. I scroll up. I overshoot into yesterday. I scroll down. Back and forth until I find today."
4. The author's comparison: "A feed opens at the newest item. A chat doesn't."

OpenAI's own Tasks docs describe the same shape of job ("summarize five industry stories every morning and include links") on the Academy Tasks page. The author's thread does not name which outlets the task reads or paste the prompt.

The rest of the four-post thread is a product critique (settings that disappear into chat), not more setup steps.

Good to know

• Results are self-reported. Profile showed 0 followers on our Nov 6 read; the account still has older posts on other topics (Apple hardware, Python), so the "every morning" claim is not a one-post spam account. Neutral pronouns used.
• The source does not name the news sites, the model or the exact clock time.
• Expect the scroll hunt the author describes, or open Scheduled and read the latest run there (ChatGPT Learn).
• A separate user (@298gama, late Nov 2 ET) also said they use ChatGPT scheduled tasks "to follow policy news every morning," which matches the pattern but is not this author's setup.
• Disclosure: the thread does not sell a template. It is an argument that AI needs a dashboard.
• Prior ChatGPT scheduled-task workflows on main cover weekly NDA filing, invoice checks and health reports, not a morning news RSS stand-in.

Set it up in Grok Bot

Same idea as a Grok Bot weekday routine that writes a short brief into chat or a file (ours, untested). Routines spend usage each run.

Starter prompt (ours, untested):

Every weekday at 7:30 AM, gather overnight news for [topics].
Write a short brief with 5 items. For each item: one-line headline,
why it matters in one sentence, and the source URL. Prefer primary
sources. If you cannot open a source, skip it and say so. Do not
post anywhere.

Last checked: Nov 6, 2026, 7:11 to 7:19 AM ET. What changed: first published Nov 6.

A T3 Code task writes the day plan into Notion before you wake (Other · reported)

A 7 AM T3 Code task writes the day plan into Notion; the author now runs ten timed tasks.

You can have a scheduled coding agent write your day plan into Notion before you open the laptop. Aryan Bhardwaj (@bhardwaj03aryan) posted on Nov 6 that "woke up and my day plan was already in notion. a t3 code scheduled task writes it at 7."

What it replaces

Sitting down first thing and building the day plan by hand. The author framed the wider habit as accidental automation: "didnt plan to automate my whole day but here we are i think."

How it works

T3 Code is Theo's multi-harness coding app (Claude Code, Codex and others behind one UI). Its scheduled tasks can run once or on a recurring schedule (scheduled tasks since 0.1.0, and one-time schedules since 0.1.7).

1. A T3 Code scheduled task runs at 7 and writes a day plan into Notion.
2. Other tasks on the same account fill the day: "a network run at 11, nightly close at 23:30, log cleanup every 6h."
3. The author counted "8 scheduled tasks" in a reply to @theo earlier the same morning, then "10 of them now" after waking up to the Notion page.
4. In that reply the author also said "opus plans, haiku subagents do the browsing and file stuff," and "the scheduled tasks are the part i use most honestly."

The posts do not paste the day-plan prompt or show a Notion screenshot. The concrete result they state is the plan already present at wake-up, plus a growing list of timed runs.

Good to know

• Results are self-reported. Account age Dec 2022, about 41 followers on our Nov 6 read. Neutral pronouns used.
• The day-plan post does not name which model wrote the Notion page. The earlier reply names Opus and Haiku for the wider factory, not for that one task alone.
• Scheduled tasks have been in T3 Code since 0.1.0, and one-time schedules since 0.1.7 (Flavio's tour).
• Disclosure: none of the author's posts here pitch a paid template or mailing list.
• Prior coverage: no T3 Code scheduled-task workflow on main; Hermes and Grok Bot morning plans are different tools.

Set it up in Grok Bot

Same shape on Grok Bot if you want Notion without T3 Code: a weekday routine plus a Notion connector (ours, untested). Cursor's routines page notes each run spends usage.

Starter prompt (ours, untested):

Every weekday at 7 AM, write a short day plan for me and write it
to the Notion page named [Day plan]. Use today's date as the heading.
Include: top 3 priorities, meetings I already told you about, and one
"if time" item. If Notion is unreachable, leave the plan in this chat
and say so. Do not email or message anyone.

Last checked: 7:24 to 7:25 AM ET Nov 6. What changed: dropped the unverified claim that T3 Code schedules follow the server's clock or time zone; scaled and centered the cover so the tent calendar stays whole in the 1:1 crop.

Opinion · Friday wrap

Build a Flare-draft, Sunburst-final GIF brand kit (Opinion · reported)

Opinion: lock brand refs, draft motion on Flare at low complexity, then finalize looping GIFs on Sunburst. Sign-out receipts and a Hark usage log rank next.

This is The Daily Bot's Friday opinion column. The picks and the ranking are ours. Every fact in it is linked.

This week's three picks share one idea: know what stayed active, or what consumed the limit, before you call the job done. OpenAI Desk wants a draft lane and a final lane for brand GIFs. Anthropic Desk wants a containment receipt after a Claude Enterprise sign-out. Hark Watch wants a log of what ate the allowance.

Our ranking: build the Flare-draft, Sunburst-final GIF brand kit first. You can start from shipped docs on Monday. The Claude sign-out checklist is second, buildable if you have Enterprise. The Hark usage log is third, because only Hark can ship it.

1. Our pick: a Flare-draft, Sunburst-final GIF brand kit

Why this week

OpenAI shipped Images 2.6 on Nov 4 with GIFs, reusable references, and GPT-Image-2.6 Flare and Sunburst. Per the announcement: "When quality is set to low, Flare can generate images in as little as six seconds." The same page says Sunburst brings "higher-quality image generation and even greater image consistency across edits when quality is set to medium or high."

The gap

The controls already exist. ChatGPT can save a reference and call it with @Name, per the Help Center. OpenAI's Nov 4 post said Save once. Create again and again.. The image generation guide already has complexity, markup edits, and GIF output on generations and edits. Per the Flare and Sunburst pages, token pricing matches: text input $5/1M, cached text $1.25/1M; image input $8/1M, cached image $2/1M; image output $30/1M.

Our read: what is missing is a packaged lane. Nothing locks a brand pack, drafts motion on Flare at low complexity, then finalizes a looping GIF on Sunburst with a markup pass. Same token rates mean teams can skip Flare and burn Sunburst latency on every throwaway pose.

What we'd build

1. Lock the pack. Save logo, mascot, and product stills once (ChatGPT @Name per the Help Center; API vision file IDs) plus one style card.
2. Draft on Flare. Rough stills or GIF motion at quality: "low" and complexity: "low" on gpt-image-2.6-flare. Per the announcement: "For most editing tasks, we recommend low complexity for faster edits."
3. Markup keepers. Transparent annotation layer only; per the guide: "Do not combine markup with mask or prompt_rewriter."
4. Finalize on Sunburst. Approved motion to gpt-image-2.6-sunburst with output_format: "gif", still refs on edits, and set frame_count / frame_duration_ms (guide defaults 9 and 200 ms).

Copyable prompt (ours, untested):

Build a small Images 2.6 brand-kit runner. Store vision file IDs for logo, mascot, and product plus a style card. Commands: draft (gpt-image-2.6-flare, quality low, complexity low), markup-edit (transparent PNG overlay, no mask), finalize-gif (gpt-image-2.6-sunburst, output_format gif, frame_count and frame_duration_ms set, still refs on edits). Refuse finalize unless a draft is marked approved. Log model, quality, complexity, and token usage per call.

Gotchas

• Omit complexity and both models use high, per the guide. "Adjusting complexity has no separate fee."
• Per the Help Center: "ChatGPT can't edit an animated GIF you upload." Keep still masters and re-animate from those.
• frame_count (2 to 16) and frame_duration_ms (20 to 2000) are GIF-only, per the guide.

What we'd want from OpenAI

A first-party brand-kit object (named refs plus a draft/final model lane). Observable signal: the Help Center or image guide gains a "brand kit" or "draft vs final" section, not only separate Flare and Sunburst pages.

2. Ranked second: a Claude sign-out containment receipt

This is Anthropic Desk's build-next pick, folded here. We have not called the sign-out endpoint ourselves.

Anthropic's November 5 release notes added a beta Admin API endpoint that signs an Enterprise member out without removing them. The guide requires write:members and the admin-user-sign-out-2026-10-08 beta header. It ends account-wide sign-in sessions and trusted devices, but only the OAuth tokens issued for the organization making the request.

The documented exclusions are the reason for the pick: API keys remain valid, Claude Code cloud sessions keep running, scheduled and GitHub-triggered routines keep starting sessions, and Cowork-linked computers stay linked. Account-level OAuth tokens not tied to an organization also survive.

Our read: build one incident record around the call, not an automatic kill switch. Confirm the member ID and affected organizations first (each linked organization needs its own key). Record sign-out separately from identity-provider containment; Anthropic's lost-device workflow starts at the identity provider. Create separate review rows for surviving keys, cloud work, routines, linked computers and account-level tokens, attach the sign-out event (org_user_trusted_devices_revoked; partial requests can show completed: false), and leave unresolved surfaces open.

Do not treat a 200 as full containment. Anthropic's response table says a 503 means "What the request already ended stays ended," and callers should send the request again. A 200 does not confirm clearing the two-step-verification skip record or removing push-notification registrations. The primary owner cannot be signed out, and a sign-in already underway can finish up to five minutes after the request.

Our read: show these as separate checks, not one green badge. Ranked below the OpenAI kit because you need Claude Enterprise and an Admin key to build it on Monday.

3. Ranked third: a Hark usage log

This is Hark Watch's build-next pick. We haven't used Hark ourselves.

On Nov 2, lyndon wrote that on Hark Pro2, "after a few requests this morning my daily limit was hit." On Nov 4, Robert Kemp wrote that "after it was syncing my Google data, I asked 2 quick questions and ran out of tokens." Adcock replied "Will take a look - you are likely outside one of our target countries", then "try again." Kemp later wrote "It's working great, thanks." Adcock did not say what he changed.

Two posts point at hidden causes. Terry Schussler told a locked-out user that Scheduled Tasks and Panels "refreshing often" will "count towards your usage limits." The user he was answering, Xstudy, had asked Hark: "Please check my account's usage log for Oct 31" / "Nov 1, and tell me what actually consumed my quota." Xstudy also asked Hark to tell users in advance if free-plan limits are being reduced.

Hark's Using Hark page says the free plan has the full feature set but "you may reach a usage cap," and that Hark Pro² "offers 2x more usage for $20/mo" and Hark Pro³ "10x more usage for $100/mo." Adcock said on Oct 30 that Hark "started limiting tokens for free accounts in some countries such as Singapore." We could not find a Hark page that lists those countries, gives free-plan amounts, or breaks usage down by task.

Our read: a percentage is not enough when the limit hits early. We'd want a usage log, one line for each chat request, Handoff session, Panel refresh, Scheduled Task run, or first-time account sync, with when it ran, what share of the daily and monthly limit it took, and whether you started it or Hark did. Only Hark can ship that surface, so it ranks third.

Why it matters at work

WHY IT MATTERS AT WORK · score 4/10
Saves: draft rounds that currently hit Sunburst on throwaway poses (our estimate: about 15 to 60 min a week once a team is shipping GIFs weekly)
For: marketing, design, brand
Setup: about an afternoon · connects: OpenAI API key, stored vision file IDs

The Claude receipt and the Hark log are asks for their vendors on this ranking, so they do not score yet. The brand-kit runner is the build you can run on Monday.

What to watch next week

• A Help Center or image-guide section that names a brand kit or a draft vs final lane (Help Center, image generation guide).
• Per-surface completion fields on Claude's sign-out docs, not only the beta label disappearing (user management).
• Any Hark page or founder post that lists what consumed a user's limit, or names the country list Adcock referenced (Using Hark, Oct 30 post).
• Cursor Live Apps beyond the Nov 5 changelog entry. We have not tried it hands-on (changelog). Grok Bot remains on v0.78.1 (x.ai changelog).

Last checked: Nov 6, 2026, 6:25 to 6:28 AM ET. What changed: editor gate fixed Xstudy quote to note_tweet wording as split fragments (source en dash between dates not copied), removed the unpublished vercel /changelog/openai-images-2-6 link, labeled the 15 to 60 min savings as our estimate, and moved connective punctuation outside the Adcock quote; re-opened OpenAI, Claude, Hark, x.ai, Cursor and the cited X posts.

Sources

• OpenAI: Introducing Images 2.6 (Nov 4, 2026)
• OpenAI API: Image generation guide
• OpenAI API: GPT-Image-2.6 Flare
• OpenAI API: GPT-Image-2.6 Sunburst
• OpenAI Help Center: Images in ChatGPT
• OpenAI on X: Save once. Create again and again. (Nov 4, 6:35 PM ET)
• Claude Platform release notes, November 5, 2026
• User management, sign a member out everywhere
• Hark: Using Hark Pro
• Xstudy (@XstudyNet), Oct 31, 9:27 PM ET
• Terry Schussler (@schussler), Nov 1, 11:14 AM ET
• Brett Adcock (@adcock_brett), Oct 30, 1:19 PM ET
• lyndon (@coffeeintoront), Nov 2, 1:35 PM ET
• Robert Kemp and Brett Adcock thread, Nov 4 evening ET
• x.ai Grok Bot changelog (still v0.78.1), checked Nov 6, 6:20 AM ET
• Cursor changelog: Build Live Apps, checked Nov 6, 6:20 AM ET

Week in review: Nov 2 to 5

Features before fixes, from Monday's and Tuesday's issues plus Wednesday's profile backlog. Thursday and Wednesday had no separate daily issue on main; Friday carries the rest. Full Tibo profile was in the Nov 5 email.

Grok Bot and Cursor

• Phone calls moved from tip to official. Monday's tip showed a user having Grok Bot call Progressive (story). Wednesday's @bot post made outside calls official (today's lead follow-on). (confirmed)
• A shared Data Bot for PMs and marketers. SpaceXSI's data team put metric rules on one Team Bot (story). (reported)
• Weekly usage meter jump. Some meters roughly doubled over the Nov 1 weekend, then partly recovered (story). (reported)

Anthropic

• Claude API message threads hit public beta, so long agent loops can send only new messages each turn (story). (confirmed)
• FlashProve lets a plain CPU check that a reply came from committed weights (story). (confirmed)

OpenAI

• ChatGPT and Codex will share one usage allowance from December (story). (confirmed)
• Profile: How Codex's Tibo Sottiaux runs Codex in public. One of Codex's most public faces posts shipping news, usage fixes and resets almost daily, and his posts show how to spend an allowance well (story). (confirmed)

Hark

• A user ran three real tasks through Hark's Mac app (story). (reported)
• Free-limit reset dates. Users say the Nov 1 reset moved their monthly limit to Dec 1 (story). (reported)

The digest

Grok Bot and Cursor

• Computer use is broken for some regions; other features are fine. Do not Reset. At 6:49 AM ET Nov 6, status.x.ai opened a minor investigating incident: "Computer use in Grok Bot fails for users in a few regions." Status: "failing for users in some regions"; "Other Grok Bot features are working normally." Users on forum 177456 and 177470 say computerUse exits in seconds with no transcript. Staff mohitjain wrote at 4:43 AM ET: "this is on our side, not your Bot's computer or anything in your setup." "No need to Reset or recreate your computer (that won't fix it and can lose unsynced work)." Colin wrote at 4:32 AM ET: "I've shared it with the team, since it appears to affect a few users right now." (confirmed)

• Bot browser tunnel error has a fix: run Update Bot's Computer. Since early Thursday, users reported that the Bot's browser failed on every site with ERR_TUNNEL_CONNECTION_FAILED when Route traffic through this computer was off. At 11:28 AM ET Cursor staff Colin wrote: "Hi all, a fix is out now." "There's no need to Reset." Open Settings, then Computer, click Update Bot's Computer, then ask your Bot to open any site. Update, Nov 6: Some users still report ERR_TUNNEL_CONNECTION_FAILED after Update (forum thread 177377 overnight posts); no new staff reply after Colin's Nov 5 fix post. Colin's Update Bot's Computer steps still stand. (confirmed, staff)

• Vault secure fill failing with driver_unavailable; sign in by screen handoff. A user reported that secure fill returned driver_unavailable for a saved Vault login. At 11:29 AM ET Cursor staff Colin replied "your saved logins are fine. The fill is failing because the bot isn't receiving a response from its browser driver when it tries to fill." He added "Other users are seeing the same thing" and "For now, the screen handoff is the reliable way to sign in." No fix or ETA yet. (confirmed, staff; open)

• Some Bots still can't get a screen: "All screens on the shared computer are in use." In a forum thread, users say every Bot's desktop and browser fails with that message. Cursor staff Colin called it "a known issue the team is aware of," said "there isn't a safe way to release the screens from your side right now," and said "A Reset isn't expected to clear this either." No staff update since. (confirmed, staff; open)

• Grok Bot computer hang from a startup-file background job. A line in ~/.bashrc or ~/.profile that starts a background program can leave Shell, Screenshot, and most /workspace reads timing out for 12 to 14 minutes while connectors still work. Reset and Update restore the same startup file, so they do not clear it. Colin wrote at 12:41 PM ET Nov 5: "The team is aware of this and is working on a fix." Workaround from that post: from the desktop app, set Execution on Local Computer to anything except Never; ask a bot to copy /home/box/.bashrc without using Shell; comment out the autostart line; copy the file back; click Update Bot's Computer (not Reset). (confirmed, staff; one public reporter)

• Scheduled Outlook webmail can work while chat refuses the browser; staff say that split is not intended. Staff mohitjain wrote at 4:40 AM ET: "This isn't anything in your setup, and there's no intended split between scheduled and chat, or between reading and sending." "Chat recently started pushing toward the Outlook connector instead of the browser, which shouldn't happen when the connector isn't usable on your account." Workaround: "Have your hourly routine do the drafting and leave it in Drafts for you to review and send (the scheduled side still works)." Or get IT to approve the Outlook connector. (confirmed, staff)

• On iOS 18, Grok Bot 1.24.1 can quit the moment you start typing. Use the mic or the desktop app. Staff mohitjain wrote at 4:16 AM ET: "This isn't caused by anything on your phone, your keyboard, or your setup, so there's nothing more you need to reinstall or reset." "Use the mic button in the composer to dictate your message." Desktop at https://x.ai/bot is unaffected. He also noted on forum thread 177488 that freezes on 1.18/1.19 when sending or returning from background are already tracked, and that 1.24.1 on iOS 18 can close when typing starts. (confirmed, staff)

• Composer 2.5 requests had issues for 50 minutes Thursday, now resolved. Cursor's status page opened a minor incident, "composer-2.5 degraded", at 10:06 AM ET and marked it resolved at 10:56 AM ET. Nothing to do now. (confirmed, status page)

• Grok Bot had a short burst of errors Wednesday evening, resolved. The status page logged a minor Degraded Performance incident for Grok Bot from 7:34 PM ET to 8:49 PM ET on Nov 4. (confirmed, status page)

• Notion pitches one shared memory for all your agents, and names Bot. Notion posted that "ChatGPT, Claude, Muse, and Bot can save what they learn to the same memory and use the same skills in Notion." Its setup guide's first step is connecting Notion's MCP server. Notion's post doesn't tag the @bot account, and we have not run the setup. (confirmed, Notion post; untested)

Anthropic

• Claude Code background subagents use less of the limit. @ClaudeDevs said at 2:29 PM ET Nov 4 that a running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less. A follow-up says sessions with background subagents used about 5% less of the limit (about 6% more prompts), that it is live for most Pro, Max and Ultra users and for everyone on Team and Enterprise, and to update to v2.1.316+. (confirmed)

• Anthropic's biology lab reports a vault-particle structure. Anthropic's biology lab says in Investigating vault particles with Claude that it determined the first 3D structure of a prokaryotic vault particle: 96 copies of a bacteriophage protein, made in E. coli in the lab. Anthropic has not yet tested whether these vaults form in the bacteria and phages where they occur naturally, and vaults' function remains unknown. By Anthropic's account, a three-day purification took about 90 minutes, and Claude, running in an internal multi-agent harness on internal compute, had a first structural model 34 hours later; one Claude protocol error cost two weeks. The preprint was submitted to bioRxiv and is not yet linked. @AnthropicAI posted at 3:11 PM ET Nov 4. (early research)

• Claude Code 2.1.318 on latest; 2.1.319 on next only. Claude Code 2.1.318 is npm latest (published Nov 3 at 7:01 PM ET). It adds the date a stored claude.ai sign-in ends to claude auth status and a stderr warning on claude -p starting 10 days before expiry, plus an effort option on a mod's $.model.fork for low-effort side jobs. The npm registry lists 2.1.319 as next (published Nov 5 at 7:58 PM ET). The official changelog has no 2.1.319 section. (confirmed)

OpenAI

• OpenAI will acquire Pierre Computer Company (deal not closed). On Wednesday, OpenAI announced it will acquire Pierre Computer Company, Inc., "the company behind Code Storage and a family of standout open-source developer tools." If you've "browsed files or reviewed a change in the Codex app, you've already used Pierre's Trees and Diffs libraries." OpenAI says Pierre's open-source projects "will remain open-source," and "Customers can continue using Code Storage after the acquisition closes while we work to bring its features and functionality into Codex." Closing is "subject to customary closing conditions, including receipt of regulatory approvals." Until then, "OpenAI and Pierre will remain separate and independent companies." No price or close date. Our read: watch for post-close Codex product notes, not a same-day capability change. (confirmed, announced)

• Daybreak Red (GPT-5.6 Cyber) reaches AWS GovCloud (US-East). OpenAI's Wednesday product post says "GPT-5.6 Cyber is now available to approved customers in AWS GovCloud (US-East) through Daybreak Red." Access "requires OpenAI Daybreak Red approval and AWS provisioning", and "Daybreak Red approval does not automatically enable model access on AWS." This is GPT-5.6 Cyber on Bedrock, not the Nov 5 GPT-6 Cyber API and Codex launch covered above. (confirmed)

• Update: ChatGPT Work and Dots errors resolved. OpenAI's status page marked the "ChatGPT Work and Dots Errors" incident resolved at 11:31 PM ET Nov 5, about 13 hours after it opened at 10:25 AM ET Nov 5. The final note says: "All affected services are operating normally again." The status page reported all systems operational when we checked at 6:03 AM ET Nov 6. (confirmed)

Known issues: - None open on status.openai.com after the Work and Dots incident resolved. Checked with the morning desk.

Hark

• iOS 2.8.0 keeps unsent drafts; Panels get clearer labels. Apple's Hark Pro listing shows version 2.8.0, released Nov 5. Release notes: "Hark now retains your unsent drafts even if the app crashes or is force quit. Panels have improved labeling." The listing shows 2,623 ratings (average about 4.71). We did not install the build. (store listing)

• Mac installer file updated Nov 5; still no release notes. The Hark.dmg linked from hark.com/mac now has a server Last-Modified of 1:34 PM ET Thursday, Nov 5. Hark has not posted release notes, so we cannot say whether it includes the Mac crash fix Brett Adcock said Nov 1 would be "in the next update this week." (installer headers, founder post)

• A user says Hark booked a go-kart day off for their son. Casey Campbell (@_campbell_casey) wrote at 10:52 PM ET Thursday: "just planned my son's day off before I'd even noticed he had one. It caught the no-school day on my calendar, remembered how much he loved driving his own kart last time, found a junior go-kart track with times and prices, and booked the tickets." (user report) Hark replied "VROOM VROOM" (company post). Confirm before spend still applies to ticket bookings.

• A user says Hark ordered a Cuisinart replacement part. Jeffrey Weichsel (@jeffreyweichsel) quoted what he says Hark wrote about matching a Cuisinart part, comparing prices, adding a $5 spatula for free shipping, placing the order, and emailing his wife. (user report quoting Hark) Hark replied. Confirm before spend still applies.

• Pro2 burn and an App Store Pro4 line. Nova (@FanPiny) wrote "Pro2 barely made it a week... 94% of the monthly limit gone, 23 days to go." (user report) Brett Adcock posted "Apple 2.0" quoting a "Hark will be the new Apple" prediction (founder post; not a launch). Apple's listing shows Hark Pro4 at $259.99; the web pricing page still lists Free / Pro2 $20 / Pro3 $100 with no Pro4. (user report, founder post, store listing, pricing page)

Known issues and workarounds

• Computer use fails for users in some regions (Grok Bot, confirmed, staff, opened Nov 6): other features fine; do not Reset or recreate your computer. Status page investigating; staff say it is on their side. Forum 177456
• Vault secure fill failing with driver_unavailable (Grok Bot, confirmed, staff; open): sign in by screen handoff instead. No fix or ETA yet. Forum 177406
• All screens on the shared computer are in use (Grok Bot, confirmed, staff; open): no safe way to release screens from your side; Reset not expected to clear it. Forum 177180
• Computer hang from a startup-file background job (Grok Bot, confirmed, staff): a line in ~/.bashrc or ~/.profile that starts a background program can leave Shell and Screenshot timing out 12 to 14 minutes. Workaround: comment out the autostart line via desktop Execution on Local Computer, then Update Bot's Computer (not Reset). Forum 177421
• Grok Bot 1.24.1 on iOS 18 can quit when typing starts (Grok Bot, confirmed, staff): use the mic button or the desktop app; nothing to reinstall. Forum 177462
• Bot browser tunnel error still reported after Update (Grok Bot, confirmed, staff fix Nov 5; some reports remain): open Settings → Computer → Update Bot's Computer, then ask the Bot to open any site. Some users still report ERR_TUNNEL_CONNECTION_FAILED overnight. Forum 177377

Coming up

On the calendar

• Monday, Nov 9: Grok Bot for sales ops playbook (planned). SpaceXSI's sales team described splitting its deal cycle across shared Bots. Held for Monday's issue. (planned; source is the company post)
• Sunday, Nov 22: Anthropic Residency applications close at 11:59 PM PT (announcement, US posting). (confirmed)
• December: ChatGPT and Codex shared usage starts. One token allowance for chats and Codex work. (confirmed, month only)

Confirmed, no date

• Computer use failing in some regions (no fix date). Status.x.ai minor investigating incident opened 6:49 AM ET Nov 6. Staff say do not Reset. Watch the status page and forum 177456. (confirmed open; fix date unverified)
• Claude Personal Finance gradual US rollout. Help Center says you might not see it yet (Help Center). (company docs; date unverified)
• GPT-6 Cyber expanding beyond the first Daybreak Red cohort. OpenAI says availability expands over time (announcement). (company post; date unverified)

Early signals (unverified)

• Hark Mac release notes or a confirmed crash-fix build. The Nov 5 DMG has no release notes, and Adcock's "next update this week" claim still lacks a matching note (founder post). (unverified)
• Pierre acquisition close. OpenAI says closing needs regulatory approvals (announcement). (confirmed announced; close date unverified)

Daily Bot Check

1. What can a Cursor cloud agent do with live apps today?
2. Who can use Claude Personal Finance in the US beta?
3. What does Codex CLI 0.167.0 open without a daemon?

Answers: 1. Build, host and publish an app, then give you a share link. 2. Pro, Max and Ultra plans. 3. The agent command center with unread marks.

You're getting this because you subscribed to The Daily Bot. Reply to this email with corrections or tips; a person reads every reply. The website is temporarily offline; fuller articles will keep arriving in this email until it is back.

How we label: confirmed means we checked a primary source or a staff post; reported means a credible claim we haven't verified; "Our read" is our own analysis.