Claude Code 2.1.323 Stops "Don't Ask Again" From Saving Secrets

AI Changelog · Numéro 32Week 46 · Nov 9–15, 2026

This week's risk sat in files and services developers rarely check: Apiiro reportedly found passwords in 1,122 public GitHub copies of Claude Code's settings.local.json, and Fireworks AI's breach reached stored secrets for some accounts and Notion AI transcripts from under 1% of workspaces. Claude Code shipped its fixes, and a loosening of managed mod controls, as ordinary changelog lines with no CVE or advisory we could find, so the changelog was the only notice.

112 announcements scanned · 28 mattered · 0 verdicts changed

À traiter avant le prochain numéro

Upgrade Claude Code to 2.1.323 and clean settings files already pushed

Claude Code 2.1.323 (12 November) makes a "don't ask again" allow rule that looks like it holds a secret, or would land in a git-tracked settings.local.json, last for the session only. Upgrading stops new leaks; it does not remove files already pushed, so check your repos and their history and rotate what those files held.

Détails et étapes

Check .claude/settings.local.json in your public repos

Claude Code saves "Yes, and don't ask again" approvals, secrets included, in .claude/settings.local.json, which it keeps out of Git only through your global git excludes. Apiiro reportedly found passwords in 1,122 public GitHub copies of that file. Untrack it, add it to the repo's .gitignore and rotate every credential it ever held.

Détails et étapes

Review deny rules before rolling out Claude Code 2.1.321

From 2.1.321 (9 November), mods Claude writes turn on without asking in auto and bypass modes, even where managed settings set allowUnsandboxedCommands: false or allowManagedDomainsOnly. Their own tool calls now skip your ask rules, so deny rules and organization policy are the controls left.

Détails et étapes

Treat Claude Code 2.1.319 and 2.1.321 as security releases

Claude Code 2.1.319 (6 November) and 2.1.321 (9 November) close a worktree sandbox escape, several Bash permission bypasses and an MCP OAuth bug that sent refresh tokens to a redirect host. There is no CVE or advisory, so the changelog is the only notice.

Détails et étapes

Rotate secrets stored with Fireworks AI and credentials pasted into Notion AI

Fireworks AI says attackers who reached its development machines on 24 October hit account and billing information and, for a limited set of accounts, secrets, datasets and incomplete models. Notion says Notion AI chat transcripts from fewer than 1% of workspaces were exposed through Fireworks and tells teams to rotate any credentials pasted into Notion AI.

Détails et étapes

Decide on Claude Data before 19 November

On Enterprise plans, Claude Data turns on by default on 19 November unless an admin changes it in Organization settings > Artifacts. Team plans already get Claude Data, Whiteboards and Sheets on by default.

Détails et étapes

Budget for Gemini 4 Carbon's price rise on 1 December

Google made Gemini 4 Carbon (gemini-4-carbon) generally available to every paying Gemini API developer on 10 November, at $2 input and $10 output per 1M tokens through 30 November and $4/$20 from 1 December. Carbon for cyber work still sits behind the Fairwind Program, and every benchmark is Google's own.

Détails et étapes

Pin claude-fable-5-1 if a workflow depends on it

Claude Fable 5.2 lists at $8/$40 per 1M tokens against $10/$50 for 5.1, and Claude Code 2.1.321 makes it the default Fable and moves saved Fable 5.1 picks to it. Every benchmark table in the launch post is Anthropic's own.

Détails et étapes

Move gpt-image-2 calls to GPT-Image-2.6 before 10 February 2027

OpenAI deprecated gpt-image-2 and its gpt-image-2-2026-04-21 snapshot, with shutdown on 10 February 2027 and gpt-image-2.6-sunburst or gpt-image-2.6-flare as the replacement. Three older image models still shut down first, on 1 December 2026.

Détails et étapes

Sur notre radar

  • In review, not yet published

    None of these drafts has passed our source review, so we report no figures from them yet: Gemini 4 Hydrogen, Meta's Muse Ember, the Gemini Decisions API, self-hosted Claude Managed Agents, Claude Meetings, the Claude Dispatch browser extension report, human review of Gemini Connected Apps data, outbound SIP calling in OpenAI's Live API, the Gemini Embedding 2 Preview route, Claroc's report on Grok-built apps, Railway's stolen AI provider keys, the TypoPrompting lookalike domains, ElevenLabs Scribe v3, VulnCheck's Glasswing tally and outbound calls in Meta's Muse agent.

  • OpenAI models in Cursor: 12 November passed with no confirmed cutoff

    OpenAI's help page, read in full on 13 November, still calls 12 November a proposed transition date, and Cursor's models page that day still listed GPT-5.6 Sol, Terra and Luna. A listing is not proof that requests succeed. Our story on it is still in review.

  • A carried-over Claude Code draft still failing review

    Our story on Claude Code putting a user's account email in the User-Agent header failed source review twice on 10 November. It stays a draft until fixed.

  • Ten verification drafts waiting for review

    The 12 November tool verification pass drafted 10 posts, among them Cursor's notice on the malicious Claude Code extension and updates to Claude's pricing page and Grok Build CLI. None has passed source review yet.

  • Claude Opus 5.5 reports with no Anthropic statement

    User reports that Opus 5.5's thinking changed on 3 November rest on Hacker News titles and an X post we could not read. A Claude Blog card saying Opus 5.5 is the default on Pro, Max, Ultra and Team plans carries no date. Held until Anthropic states a change.

  • Promised weights and previews with no primary source

    Thinking Machines' Inkling 2 Small weights, promised for "next week" on 5 November, were still missing on 13 November. A Qwen4-Max preview, Edge0-145B and a rumoured Apple Mac app for local models rest on X posts, Hacker News titles or reports. Held until a vendor page names access, licence or price.

  • Security research with no affected directory tool

    A Firecracker VM escape write-up that depends on an earlier host-memory primitive and runs without the jailer, an SGLang flaw used against NEAR AI with no CVE or fixed version named, and a Microsoft research result on backdooring agents by editing fewer than 10 neurons. Held until an advisory or a directory vendor names an affected product.

  • Single reports about ChatGPT surfaces

    A claim that ChatGPT Desktop forwards your SSH agent to remote machines cites a pull request that was closed unmerged, and a developer's report that the Apps SDK routed a tool call through the wrong connector has no OpenAI reply. Held until OpenAI documents or changes the behaviour.

  • Policy and enterprise items held at headline level

    New York's RAISE Act, where the headline and body we read disagree on when it takes effect; the Enterprise Frontier Safeguards beta that came with Claude Fable 5.2; and an Amazon story on replacing Sonnet 4.6 with Strands Decider, read only as a Hacker News title. Nothing read changes a reader's setup yet.

Écarté

  • Re-reports of stories already covered: Gemini Omni 1.2 Flash, Perplexity Teammates, GPT-6 Cyber and the OpenAI Linux privilege escalation from our week 45 issue, plus a benchmark of Fable 5.2 against GPT-6 Cyber in The Hacker News.
  • Funding, business and consumer news from TechCrunch and The Verge, including the Nvidia and Jensen Huang coverage, Anthropic IPO stories, OpenAI's $10B quarter and Apple Intelligence Usage Packs. None moves a tool, model or price for developers.
  • OpenAI customer, program and research posts, such as ChatGPT Enterprise in AWS Marketplace, BioBasicsBench, ChatGPT Plus for veterans and two alignment research posts. No API or plan change.
  • General security items from The Hacker News, including Microsoft's 1,259 patches, a Chrome zero-day, SonicWall, Citrix and a Git recursive-clone flaw. No AI tool or model in our directory, and the Git fix names no coding agent.
  • Goldman Sachs restricting Claude Code and GitHub to engineers, reported by State of AI citing Bloomberg. One employer's internal policy, and we did not read Bloomberg.
  • Release notes with no default, security or price change: Codex 0.167.0 to 0.167.2, Claude Code 2.1.322 (fast mode on Vertex), Cursor's voice chat with agents and GitHub admin entries such as Copilot seat management and Dependabot pausing on inactive repos.
  • Benchmark, customer and vertical posts: Artificial Analysis' Decision Index v2, Claude Blog customer stories such as J. Crew, Curative and JPMorganChase, the Claude for Manufacturing plugin and LiteLLM's day-0 model posts.
  • Index and changelog pages with nothing new: the xAI newsroom, DeepSeek's change log, Mistral news, Cloudflare's AI tag, the NVIDIA blog and Google's Models and Research index, which loaded only its navigation for a fifth run.
  • Hacker News community projects, essays and off-topic items, such as Libghostty, Rust 1.100, .NET 11, Val Town winding down and open-source projects banning AI code. No AI tool angle.

Notre avis : Our verdict on this week: upgrade Claude Code to 2.1.323, then search every repo and its history for .claude/settings.local.json and rotate whatever it held, because upgrading stops new leaks but does not clean files already pushed. If your organization relies on managed sandbox keys, review deny rules and organization policy before rolling out 2.1.321 or later, since mods Claude writes now turn on without asking in auto and bypass modes. If you stored secrets with Fireworks AI or pasted credentials into Notion AI, rotate them. No rating moved this week. We weighed three claims and kept every rating where it was: Claude Fable 5.1 and Gemini 4 Carbon were already conditional, and moving Gemini 4 Argon or Gemini 3.8 Flash would have rested on Google's own benchmarks alone. Elsewhere, Gemini 4 Carbon is now on the paid API at $2/$10 per 1M tokens until 30 November and $4/$20 from 1 December, Claude Fable 5.2 cuts the list price to $8/$40, and OpenAI shuts down gpt-image-2 on 10 February 2027. Gemini 4 Hydrogen, Muse Ember, the Gemini Decisions API and the Railway and Grok-built app exposures are still drafts in review, so their figures are not in this issue. How we count: daily scans on November 10 and 13 fetched 112 items (53 and 59). The scout kept 9 act, 3 verdict and 16 fyi stories, and logged 16 watch stories and 46 noise rows. The curator turned the 28 kept stories into 25 news drafts and rejected the 3 verdict claims. Source review on November 10 and 13 passed 9 drafts (4 and 5), all from this week's scans; a carried-over story on Claude Code's User-Agent header failed twice and stays a draft. Verification passes checked 55 models on November 9 and 45 tools on November 12, and the tool pass drafted 10 posts. Directory growth on November 10 and 13 added 5 models, 1 tool and 4 comparisons. A site audit on November 10 logged 42 findings across 39 pages. Translation backfills ran on November 10 and 13.

Neomanex Utility editorial

Recevez le numéro 33

Les changements de verdict et les échéances de la semaine prochaine, dans votre boîte mail.

En vous abonnant, vous acceptez notre Politique de confidentialité. Désabonnement à tout moment.