Cursor says a fake Claude Code 2.1.313 affected at most 5,000 users. One command tells you if you're one, plus Astra and Hark on Mac.

The Daily Bot

No. 022 · Friday, Oct 30, 2026 · Read on the site

If you installed or updated Claude Code inside Cursor on Oct 29, today's one job is a one-line check for an attacker's SSH key, and everyone should update the extension to 2.1.314. Also inside: how to handle OpenAI stopping an Astra agent, Hark's new Mac app, Claude's slower but more detailed usage API, Cursor's overnight Cloud Agents outage and two setups people run. The week in review and the Friday column lead Monday's email.

In this issue: Tampered Claude Code extension · What's new (3) · Since our last email (2) · Workflows (2) · Digest (10) · Known issues · Coming up

Today's lead

Check your machine for the tampered Claude Code extension in Cursor (Cursor · confirmed)

What happened: Cursor's security notice says a "maliciously modified version of Anthropic’s Claude Code extension", version 2.1.313, was in its extension marketplace "Between about 2:27am PT and 1:35pm PT" on Oct 29, which is 5:27 AM to 4:35 PM ET. It "was not developed by Anthropic." Per the notice, it collects host details, tries to add an attacker key labeled svc-telemetry to authorized_keys, turn on the local SSH service and open a reverse SSH tunnel, and keeps retrying "roughly every 60 seconds." Cursor says it removed the version "immediately upon detection," shut the marketplace down "out of an abundance of caution," confirmed no other extensions were tampered with, and that "The attacker server no longer functions."

Who's affected: Anyone who "installed or updated to Anthropic.claude-code 2.1.313 from Cursor VS Code extension marketplace between 2:27am PT and 1:35pm PT on October 29, 2026". Cursor says "At most 5,000 users were affected, and we will reach out directly." If you are not sure, run the check anyway.

Not affected: Cursor posted on X at 3:46 AM ET Friday that the incident "does not impact other extensions downloaded from the cursor.com/marketplace, Grok Bot, Grok or our other SpaceXSI products." Cursor has "no evidence that other versions of the extension or the official distribution are affected." OpenSourceMalware lists the npm package, the Open VSX extension and the Microsoft extension as "NOT malicious."

Why it matters: This hit developers who run an AI coding agent inside their editor, the people most likely to have cloud keys, repo tokens and production sessions on the same laptop. Removing the extension is not the whole fix: Cursor warns that "Uninstalling the extension may not be sufficient" because "the payload drops standalone persistence scripts that keep running." Turning on SSH "requires admin rights, which the malicious extension does not have on most macOS and Windows hosts," but Cursor still says "Follow the remediation steps below regardless."

What to do:

1. If you installed or updated Claude Code in Cursor on Oct 29, open a terminal on that computer.
2. On a Mac, run Cursor's check: grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted". On Windows, run the PowerShell check from Cursor's notice.
3. If the key is found, isolate the computer, run Cursor's remediation script for your platform, decide whether to rebuild it, and rotate credentials and sessions that computer could reach, from a different, trusted computer.
4. Either way, update Claude Code for VS Code to 2.1.314.

The cleanup scripts are remediate-macos.sh (run it "as the affected user, not as root") and remediate-windows.ps1 (run from PowerShell opened with Run as Administrator). Questions go to [email protected], per the notice.

Prompt (ours, untested):

On this Mac, run grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted" and tell me the exact output. Change nothing.

What to watch: Direct notices from Cursor to affected users, and how the bad build got in, which the notice does not say. Anthropic "did not immediately respond to a request for comment," CNA reported.

Read the full story

What's new

Cursor's Cloud Agents outage from 7:29 PM to 3:11 AM ET and its extension marketplace incident from 5:38 PM to 3:12 AM ET are both resolved, per the status data (details in the digest). Claude Code 2.1.314 reached npm at 8:12 PM ET Thursday; stable is still 2.1.308. status.openai.com shows no new incidents since Tuesday, checked 6:10 AM ET Friday. (confirmed)

How to handle OpenAI pausing or stopping your Astra agents (OpenAI · confirmed)

If you build agents on Astra, OpenAI's monitors can stop a conversation, and OpenAI does not routinely escalate flagged cases to its own reviewers, so your team responds. Per OpenAI's API docs, a block before streaming begins returns HTTP 403 with code misalignment_policy_violation, and streaming apps must also catch the error mid-stream. Stop that conversation's actions without retrying, and subscribe to safety alert webhooks so the alerts reach your own team (Help Center). Read the full story

Hark is on Mac now: what the new desktop app is and isn't (Hark · reported)

If you use Hark on a Mac, you can stop keeping it in a browser tab. Hark posted at 12:35 PM ET Thursday: "Hark is now available on Mac." Hark hasn't published release notes, so we downloaded the installer from hark.com/mac and read its files without running it: version 0.1.18, macOS 14 or later, Apple silicon and Intel. Download it from that page, not a search; another Mac app is also called Hark. Read the full story

Claude's new usage and cost API adds detail but lags up to 24 hours (Anthropic · confirmed)

If you pull Claude API usage or cost into a dashboard or a finance report, Anthropic's new beta Usage and Cost Analytics API breaks it down by API key, workspace, model and the person or service account behind each request (release notes). The trade is time. Per the docs, "Data typically refreshes within 4 hours of the underlying usage and can take up to 24 hours," where the legacy API's typically lands within 5 minutes. Our read: this is a reporting API, not an alarm. Read the full story

Since our last email

These went up on the site after Thursday's 8:29 AM ET email.

Cursor

• Cursor 3.28 on Windows breaks agent chat; staff say a fix is coming (Cursor · fix · confirmed). If agent chat on Windows answers every prompt with "An unexpected error occurred," chat in the classic IDE window (Ctrl+Shift+N, or File > Open IDE), or reinstall 3.27 from cursor.com/download and set Update: Mode to none. Colin from Cursor replied at 4:58 AM ET Thursday: "We’ve identified a fix and will roll it out soon." Story

Grok Bot

• Grok Bot for presales: a battlecard the morning a competitor launches (Grok Bot · playbook · reported). A Docusign solution architect's Compete Bot builds battlecards and talk tracks, and his weekday noon report stays quiet unless something matters. It's a self-report we haven't run; details are in Workflows below. Story

Workflows and Setup of the day

Both are self-reported, and we have not run either.

A weekly CRM release roundup (Grok Bot workflow). Who: Steve (@SteveChipman) at CRM Switch. Trigger: a routine every Monday morning. Rules, in their words: "Only items dated the previous week make the cut, and quiet vendors get listed as such," and "A bot pushes the information to a CRM Switch subdomain page hosted on Netlify." Result: the live page sorts items by job, labels each Live, Beta, Early access or Founder claim, and ranks "unverified vendor or founder claims" lowest. Cost: Cursor's routines page says "Each run spends usage," and a weekly routine is about four runs a month. The posts don't say how the bot deploys to Netlify.

Every Monday at 8 AM, check what these vendors shipped last week
(Monday to Sunday): [vendor list with changelog URLs].
Only include items dated last week. Group them by the job they help with,
give each a status (live, beta, early access, claim only) and a source link.
List vendors with nothing new under "Quiet this week" and say what you checked.
Send the write-up to me in this chat. Don't post or publish it anywhere.

(Starter prompt is ours, untested.) Read the full story

Setup of the day: a battlecard the morning a competitor launches (Grok Bot playbook). Who: Zach Woodward, whose X bio reads "Entrepreneurial Solution Architect @docusign solving real business problems," writing in his Oct 2 post. Tools: a chief of staff Bot that "hands work to about 15 specialist bots," one of them a Compete Bot that "tracks AI industry news and competitors, and builds battlecards and talk tracks." Result, his words: "A competitor launched an AI product for legal teams. I had a battlecard the same morning." His noon report "flags anything worth acting on. Most days it stays quiet, which is the point." He gives no time saved or deal results. Cost: he wrote "$200 a month" on X in September but doesn't name the plan. Disclosure: he builds his own products, and we found no paid tie to Cursor, xAI or SpaceXAI.

Our setup steps, built from his post and the docs (he hasn't shared his prompts):

1. One Bot for one beat, separate from your other Bots.
2. Define the artifact: agree on the battlecard sections before the first launch.
3. Run it once by hand on a launch you already know well.
4. Add the quiet report, and make "send nothing" a valid result.

Read the full story

The digest

Grok Bot and Cursor

• To stop one Bot, press its Stop button instead of typing "stop". Telling a Bot to stop can halt your other Bots too. Mohit, from Cursor staff, said at 5:13 AM ET that "when a bot is told to “stop”, it treats that as “stop all of my work”, and that includes sending a stop to the other bots it recently handed tasks to." It's "a known issue we’re already tracking." Until it's fixed, "the Stop button in that bot’s chat (shown while it’s working) stops only that bot and its own background tasks, and it doesn’t send a stop to your other bots." (confirmed, staff)
• Cloud Agents were down for some users for about eight hours, now resolved. The Cursor status page listed "Cloud Agent Degradation" for Cloud Agents, Automations and the IDE from 7:29 PM to 3:11 AM ET. One forum report said every launch had failed "with “Agent encountered an error. The agent stopped because of an unexpected error.”" since about 6:47 PM ET. Colin, a Cursor Community Support Engineer, said at 4:40 AM ET that Cloud Agents "should be running smoothly again." A separate "Auto Review Degradation" for Grok Bot ran from 7:44 to 9:20 PM ET and is resolved (status data). (confirmed)
• Cursor's extension marketplace is back. The status page's "Investigating Service Degradation with VS Code Extensions" incident ran from 5:38 PM to 3:12 AM ET and is resolved (status data). To a user getting HTTP 503 errors on extension search, Colin said at 4:25 AM ET: "The marketplace should be back." Cursor's security notice says it shut the marketplace down "out of an abundance of caution" after pulling the tampered Claude Code build (see today's lead), but the status incident doesn't name a cause. (confirmed)

Anthropic

• Claude API requests can now be 128 MB. Messages and Token Counting requests on the first-party Claude API can now be 128 MB, up from 32 MB, per the Oct 28 release notes. @ClaudeDevs said: "Image-heavy requests like long computer-use sessions and big PDFs were previously hitting 32 MB with plenty of context window left." It added that requests with more than 20 images now allow 3000 px per image, up from 2000 px. The API overview lists lower limits elsewhere: Bedrock 20 MB, Google-managed Google Cloud 30 MB, and 32 MB on Claude Platform on AWS, Anthropic-managed Google Cloud, Microsoft Foundry and the Sessions, Agents and Environments endpoints. (confirmed)
• Claude mobile apps get account switching, iPad side view and 33 languages. Anthropic listed a batch of Claude mobile app upgrades in a @claudeai thread on Thursday. On iPad and iPhone Duo, what Claude makes opens beside the chat. On iOS and Android you can stay signed into more than one account and switch without signing out. The app on web, desktop, iOS and Android adds 24 languages, 33 in all. Android on Chromebooks and foldables gets a resizable sidebar, trackpad, right-click and Claude Desktop's shortcuts. Voice mode answers sooner and works with Apple CarPlay (MacRumors reported CarPlay support Sept 4). Pull-down search is coming soon, per Anthropic. (confirmed)
• Anthropic says Claude has checked its own chats for misuse since June. Anthropic highlighted on X Thursday "self-assessment", research it first posted Oct 16: right after responding, Claude gives a short safety assessment of the conversation for offline review. The post on Anthropic's new Safeguards blog says that on Claude Fable 5 it flags 78% of conversations a stronger prompted classifier flags as potentially harmful, and 94% of high-priority ones, at a 1% false-positive rate. Anthropic says it adds less than 2% to inference compute, and has run since June. The post says: "Self-assessment collects no additional data from users." It doesn't run on zero-data-retention traffic today. Per the post, "An assessment is a signal for further review." (confirmed)
• Claude Code 2.1.314 can block writes outside your project. Claude Code 2.1.314 reached npm at 8:12 PM ET Thursday (npm). Its changelog adds permissions.blockWritesOutsideWorkingDirectories, which makes Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode. The setting isn't in the Claude Code docs yet, so this rests on the changelog alone. VS Code and Claude Desktop's Code tab gain idle compaction: a 200K+ token conversation left idle is summarized, and the summary is used if the prompt cache has expired. Fixes include sessions failing with "Invalid effort level" when an admin caps effort (Claude Code now lowers the effort and retries) and claude -p exiting 0 with no output on a failed start. Stable is still 2.1.308. (confirmed)

OpenAI

• OpenAI rewrites its Model Spec around five values. OpenAI updated the Model Spec Thursday, its "public description of how we want our models to behave." Per the post, the update "reorganizes its guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care." It "adds guidance on context, oversight, and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance," and the Spec now refers to the model "as “GPT” rather than “the assistant”." OpenAI adds: "our models do not yet fully reflect it." @OpenAI reposted the announcement from Jason Wolfe, whose X bio reads "alignment and the model spec @OpenAI". (confirmed)
• Codex CLI 0.166.0 adds codex doctor --network-only. Codex CLI 0.166.0 shipped Thursday at 3:47 PM ET. Per the release notes, it "Added codex doctor --network-only for focused configuration, connectivity, and WebSocket diagnostics," "Added codex queue --resume to resume unloaded threads after queueing a message," and "Added GPT-6.1 Terra to the bundled, Vertex AI, and Amazon Bedrock model catalogs." It also "Preserved reasoning-token accounting across resume and fork to avoid premature compaction." Update with npm install -g @openai/[email protected]. (confirmed)
• Known issues: None. status.openai.com shows no new incidents since Tuesday's GPT-6.1 Sol degraded performance incident, checked 6:10 AM ET Friday.

Hark

• Hark's Android app updates as iPhone ratings climb. The Google Play listing for Hark Pro now says it was updated Oct 29 (on Thursday it still said Oct 25), with no release notes shown. On iPhone, the App Store listing is still on version 2.6.0 and showed 4.69 stars from 2,173 ratings at 6:17 AM ET Friday; on Thursday we reported 4.67 from 2,081. Our math: 92 new ratings in a day. Hark's status page showed "All Systems Operational" with no open incidents. (store listings, status page)

Known issues and workarounds

The newest open items from Known issues:

• Tampered Claude Code 2.1.313 in Cursor's extension marketplace (Cursor, confirmed): run the one-line svc-telemetry check, clean up from Cursor's notice if the key is there, and update to 2.1.314 either way. Details
• Cursor 3.28 on Windows breaks agent chat (Cursor, confirmed, tracked): use the classic IDE window's Agent chat, or reinstall 3.27 and turn off auto-update. Details
• Bot computer updates on Oct 17, 21 and 22 dropped installed apps (Grok Bot, confirmed, tracked): keep a setup script your Bot reruns after an update; files, Bots and logins survive. Details
• Bugbot stalled on Bitbucket, and GitLab automations won't start (Cursor, confirmed, tracked): don't reinstall; start GitLab automations manually from the dashboard, and comment bugbot run on unreviewed Bitbucket PRs after the fix. Details
• Bots vanished after the X Subscription switch (Grok Bot, confirmed, tracked): don't reset, delete or pay again; the account-combine step can leave your Bots on one account and your plan on another. Details
• Bots show no usage after a half-finished X Subscription switch (Grok Bot, confirmed, tracked): don't buy a second plan; email support. Details

Coming up

What to watch from Monday, Nov 2 through Friday, Nov 6, from today's issue.

On the calendar

• Codex x TwitchCon hackathon entries close Monday, Nov 2, at 2:00 PM PT (5:00 PM ET). The hackathon FAQ says a saved draft is not entered, so click Submit (our story). (confirmed)

Confirmed, no date

• A fix for Cursor 3.28 agent chat on Windows. Staff say a fix is identified. Watch for a 3.28 patch or a staff "fixed" post (our story). (staff, date unverified)
• Pull-down search in Claude for iPhone and iPad. Anthropic says it is coming soon. (company post, date unverified)

Early signals (unverified)

• An OpenAI app simplicity and performance sprint (his claim, unverified). Andrew Ambrosino, whose X bio reads "Codex & ChatGPT @OpenAI", posted at 1:18 PM ET Thursday: "we today kicked off an intense sprint to dramatically improve the app simplicity and performance. i will oversee this. more to come." He didn't name the app or give a date. (staff claim, unverified)
• Grok Bot calls (unverified). Elon Musk says Grok Bot can call a person or business and send you a note, but the feature isn't in the changelog yet (our story). (reported, unverified)
• Hark Mac release notes or voice mode (unverified). Hark hasn't posted release notes or a help page for the Mac app, and it hasn't announced a voice mode, though the app's microphone text mentions "voice conversations" (our explainer). (unverified)
• An amended Anthropic S-1 (no date, unverified). Watch for the price range, the share count and the Oct 31 run-rate revenue figure left blank in the first filing (our story). (unverified)

Every story in this email

• Check your machine for the tampered Claude Code extension in Cursor
• How to handle OpenAI pausing or stopping your Astra agents
• Hark is on Mac now: what the new desktop app is and isn't
• Claude's new usage and cost API adds detail but lags up to 24 hours
• Cursor 3.28 on Windows breaks agent chat; staff say a fix is coming
• Grok Bot for presales: a battlecard the morning a competitor launches
• A Grok Bot that writes a weekly CRM release roundup every Monday
• Full issue, Oct 30

You're getting this because you subscribed at thedailybot.vercel.app. Reply with corrections or tips; a person reads every reply.

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.