The Diff

Archives
Log in
Subscribe
November 11, 2026

No. 039: Anthropic says it is prototyping secret scanning for Claude Code - and that auto mode is the real answer. Fable 5.2's report card: #1 on the index, $26.54 per task, one dev's $34 bill.

Anthropic answered the exposed-secrets report where it blew up - a 624-point Hacker News thread: secret scanning is being prototyped, no date, and "everyone should be using auto mode." Plus: Fable 5.2’s first independent report card, and one developer’s $34 bill. Plus the quiet ones nobody talks about.
View in browser: The Diff No. 039
The Diff  No. 039
Wednesday, 11 November 2026  ·  AI dev news, only what is verified
Anthropic announcement card for Claude Fable 5.2 and Mythos 5.2
Anthropic answered the exposed-secrets report where it blew up - a 624-point Hacker News thread: secret scanning is being prototyped, no date, and "everyone should be using auto mode." Plus: Fable 5.2’s first independent report card, and one developer’s $34 bill.

Boris Cherny, Anthropic’s head of Claude Code, responded to the Apiiro findings in a Nov 7 Hacker News thread, identified by VentureBeat and Apiiro’s own update: the team is prototyping built-in secret scanning and better auto-mode catches, with nothing shipped and no date. His standing advice: stop saving approvals, turn on auto mode.

Two days after launch, Artificial Analysis still has Fable 5.2 first on its Intelligence Index at 65 - and at $26.54 per task, its most expensive entry. One developer ran two real tickets through it and published the bill: about 42.5M tokens, roughly $34 at API prices.

A vendor answer with no shipped fix, and a report card that reads: powerful, verbose, pricey.

IN THIS ISSUE
01
🔐 Anthropic’s answer to the 1,122 exposed settings files: prototyping secret scanning - and telling everyone to switch to auto mode
Boris Cherny, Anthropic’s head of Claude Code, answered the Apiiro research in a Nov 7 Hacker News thread: no fix shipped, prototypes named, and a push to retire "Yes, and don’t ask again" approvals. Apiiro says auto mode leaks too.
📊 Fable 5.2’s first report card: still #1 on the index at $26.54 per task - and one developer’s two tickets cost about $34
02
📊 Fable 5.2’s first report card: still #1 on the index at $26.54 per task - and one developer’s two tickets cost about $34
Artificial Analysis keeps Fable 5.2 at 65 and first place, and at the top of its cost table. Zapier’s AutomationBench shows 51.4%. Adam Moore ran two real tickets through it and published the token bill.
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
01  HN thread 50407882, 624 points at 06:31 IST · VentureBeat + Apiiro’s update

🔐 Anthropic’s answer to the 1,122 exposed settings files: prototyping secret scanning - and telling everyone to switch to auto mode

Boris Cherny, Anthropic’s head of Claude Code, answered the Apiiro research in a Nov 7 Hacker News thread: no fix shipped, prototypes named, and a push to retire "Yes, and don’t ask again" approvals. Apiiro says auto mode leaks too.
SCENE

Friday night, 624 points, 427 comments: Apiiro’s post about passwords saved by "Yes, and don’t ask again" sits at the top of Hacker News, and the first reply begins "Boris from the Claude Code team here." The account is bcherny; VentureBeat’s reporting and Apiiro’s updated post identify the poster as Boris Cherny, Anthropic’s head of Claude Code. He then spent the evening in the replies, fielding one developer after another.

THE THING

What the Nov 7 comments say, verbatim where it matters: "Historically Claude Code has not done secret scrubbing (there are many good tools for that), but we are looking into what we can do to prevent this." On how: "making sure auto mode catches this, building in basic secret scanning, a few other ideas. Going to prototype a few of these to understand the tradeoffs better before we ship something." No shipped fix and no date - VentureBeat confirms none was announced. Cherny framed the approval flow itself as the problem: "the threat model is that human attention is limited," and auto mode "performs an order of magnitude better in practice." Two of his claims sit against Apiiro’s data: that the exposed files "were manually created and checked in," and that auto mode is the answer - Apiiro says it reproduced the leak with auto mode in the loop, and counted 539 repositories where a commit co-authored by Claude touched the file. It did not test whether the credentials were live. Separately, Claude Code 2.1.321 (Nov 9) fixed an MCP OAuth token redirect leak and three permission-check bugs - worth installing, and not this issue.

✂️ STEAL IT

Steal the response-reading habit: when a vendor answers in a comment thread, split the promise ("going to prototype") from the shipped thing (2.1.321, a different set of fixes). Then take the ten-minute version of his advice yourself: turn on GitHub secret scanning, and search your own repos for settings.local.json before someone else does.

Watch  VERDICT

Do the hygiene today - secret scanning on, grep your repos - because that part needs no vendor. Then watch the Claude Code changelog for the real fix: Cherny named prototypes, not a release, and Apiiro’s auto-mode caveat means "just switch modes" is not the whole answer.

HN: the thread with Cherny’s replies  ·  VentureBeat: the response  ·  Apiiro: the research (updated)  ·  Claude Code changelog

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeper Already knew it
Anthropic announcement card for Claude Fable 5.2 and Mythos 5.2
02  Artificial Analysis + Zapier, Nov 10-11 · a field report with receipts

📊 Fable 5.2’s first report card: still #1 on the index at $26.54 per task - and one developer’s two tickets cost about $34

Artificial Analysis keeps Fable 5.2 at 65 and first place, and at the top of its cost table. Zapier’s AutomationBench shows 51.4%. Adam Moore ran two real tickets through it and published the token bill.
SCENE

Adam Moore handed Fable 5.2 two real tickets on his own site - a PTO API and a page redesign - from his phone, with done criteria in the prompt and a rule to log what it got wrong. His spec asked for fields his database does not have. The model thought for 1 minute 21 seconds, decided adding the columns was "a product call I shouldn’t make on my own," built against the real schema, returned 400s naming the missing fields, and marked the PR "Refs #77" so the decision stayed open for him.

THE THING

The independent numbers, two days in: Artificial Analysis holds Fable 5.2 at 65 on its Intelligence Index, #1 overall, five points over Gemini 4 Carbon - and $26.54 per task, 44.4 minutes per task, 275M output tokens against a 72.8M median; its page calls the model "very verbose." It leads five of AA’s nine evals, including HLE-Diamond at 80%. Zapier’s AutomationBench reports 51.4%. Moore’s API ticket took about 25 minutes, 150 tool calls, +1163/-42 lines across 35 files, tests from 203 to 221, merged. The model also fixed a date-validation bug nobody had filed, with a test, and broke its own code three ways on purpose to prove the tests bit. The bill, from his Cursor usage export: about 42.5M tokens, roughly $34 at the $8/$40 API prices Anthropic’s changelog confirms; the redesign alone cost him $21.81 on-demand. One developer, two tickets, his repo.

✂️ STEAL IT

Steal Moore’s harness, not just his conclusion: done criteria up front, an honest-log rule in the prompt, and make the agent break its own code to prove the tests work. And copy his accounting - export the token counts and price the run yourself instead of trusting what the plan calls "included usage."

Try today  VERDICT

Run Moore’s test on one real ticket of yours and price it at $8/$40 before routing anything to Fable. It fails you on cost discipline: 42.5M tokens for two tickets is exactly the verbosity AA measured, landing on a personal bill.

Adam Moore: two tickets, real bill  ·  Artificial Analysis: Fable 5.2  ·  Zapier: AutomationBench report  ·  Claude Code changelog (pricing)

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
📊 THE ATTENTION GAP
HN points at 06:31 IST. Attention, not quality.
Anthropic’s HN answer on the exposed settings files
624 HN
"The SEC Should Block Anthropic’s IPO" (opinion)
406 HN
Gemini 4 Carbon general availability
281 HN
ALSO ON THE RADAR

Anthropic started its IPO roadshow. 825,000,000 Class A shares at an expected $70.00-$78.00, Nasdaq symbol "ANT," per Anthropic’s own announcement; the roadshow has not priced. The No. 037 context: Goldman pulled Claude Code from its bankers and traders weeks before taking Anthropic public. Anthropic →

Gemini 4 Carbon is generally available. Google’s claims: #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench "at more than 2X lower cost per task compared to the next best external model" - Google’s wording, no independent check seen. Artificial Analysis currently scores Carbon at 60 against Fable 5.2’s 65. Google →

Tencent’s Hy4 weights are actually downloadable now. The preview repo carries 131 safetensors shards of about 19.3 GB each - the weights whose absence No. 034 noted. Hugging Face →

Muse Ember is live as muse-ember-1.0. Meta Model API, 1,048,576-token context, $0.10 per million input and $0.50 per million output tokens up to 200K input, higher above; VentureBeat confirmed the endpoints and pricing Nov 9. VentureBeat →

🧪 ONE THING TO TRY TONIGHT
Ten minutes of secret hygiene.
Open GitHub code search and look for settings.local.json across your own public repos - and your org’s. Turn on secret scanning and push protection where they are off. Then open your Claude Code permission rules and read them once: every line is a command string you approved, and if a secret ever rode along in one, it is sitting in that file. Anthropic’s advice in the thread was the same idea in vendor clothes: auto mode on, saved approvals out.
What this issue could not see. The Hacker News comments quoted are from the account bcherny; VentureBeat’s reporting and Apiiro’s updated post identify the poster as Boris Cherny, head of Claude Code - the quotes are the account’s words, the identification is those outlets’. Apiiro’s updated research page would not render for direct reading this morning, so its contents are described via VentureBeat and Writeble; Apiiro sells security products adjacent to its findings, and it did not test whether the exposed credentials were live. Claude Code’s 2.1.321 changelog was read in full: its security fixes do not cover the settings.local.json behavior. Artificial Analysis numbers are its own measurements at Max effort. Adam Moore’s report is one developer, two tickets, his own repository, his screenshots as evidence. Google’s Gemini 4 Carbon benchmark claims are Google’s, and Anthropic’s IPO terms are Anthropic’s. VIDEO OF THE DAY is omitted for the sixth day running: no fresh, relevant and verifiable video on today’s stories - we omit rather than pad. Reddit remains absent by policy; X blocks anonymous access and was not consulted. Deduped against every published issue, No. 001-038, via the kit shipped-index and the live home archive: Anthropic’s response, the independent Fable 5.2 numbers and Moore’s field report are new; the Apiiro findings themselves were No. 036 and are referenced, not re-reported.
Talk to the editor 💬

Reply to this email with anything: what to chase, what to drop, what you want more of. Every reply shapes the next issue.

More of this kind of issueLess of this
Images: credited in each report; source screenshots and announcement artwork, not our tests.
Don't miss what's next. Subscribe to The Diff:
← Newer No. 040: Google's Gemini Decision 1 answers 1,000 questions in one pass - no text, output free. GitLab ships the settings.local.json block Anthropic hasn't. Older → No. 038: Anthropic says Fable 5.2 tops its coding benchmark - and still recommends Opus 5.5. Meta says Muse Ember matches Haiku 5.5’s best score at a third of the cost.
Powered by Buttondown, the easiest way to start and grow your newsletter.