AI Brief, November 10: Claude Fable 5.2, Meta Muse Ember, Claude Code secret leaks

Anthropic ships Claude Fable 5.2 and Mythos 5.2, Meta launches the small Muse Ember, and Claude Code settings files are leaking secrets on GitHub.

Two model launches led the day: Anthropic's Claude Fable 5.2 and Mythos 5.2 at the frontier, and Meta's Muse Ember at the cheap end for agent grunt work. Security took up the rest, from Claude Code settings files leaking secrets on GitHub to DeepMind formally verifying Signal's cryptographic code.

Anthropic releases Claude Fable 5.2 and Mythos 5.2 with faster output and a price cut

Anthropic says the two models share the same underlying model: Fable ships to all users with cybersecurity and biology safeguards, while Mythos is gated behind its Life Sciences and Cyber Verification Programs. The company says outputs are up to 30% faster than the 5.1 models, at a lower price, with an optional fast mode that speeds generation up to 2.5 times. Anthropic reports that a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, compared with 11 days for an earlier unreleased research version, and that over 90% of changes merged into its core codebase are now written by Claude with human guidance. It still recommends Opus 5.5 as the default for most coding and knowledge work.

Why it matters: Fable is positioned for open-ended, long-horizon technical work rather than daily coding, so the practical question is when to escalate to it; Simon Willison highlights Anthropic's own prompting guidance that telling the model the codebase is its own makes it far more likely to refactor repeated code instead of piling on.

Sources: anthropic.com / Simon Willison / Simon Willison / Hacker News discussion

Meta launches Muse Ember, a small model aimed at repetitive agent tasks

Meta Superintelligence Lab released Muse Ember, which it describes as its smallest and most efficient model, built for tool calls, coding subtasks, routing, extraction, and long document parsing. Meta says it matches Claude Haiku 5.5's best score at about a third of the cost and beats every GPT-6 Luna setting on Artificial Analysis, and that it sits on that leaderboard's efficiency frontier. It has a 1M-token context window, up to 512K output tokens, and native text, image, audio, and video input. TechCrunch reports it is the fourth model under the Muse brand, after Spark, Glimmer, and Wonder.

Why it matters: It gives builders another option in the Haiku and Luna tier for high-volume agent steps, with the intended pattern being a larger model that plans and Ember that executes; Simon Willison notes that per-token comparisons are getting less useful because reasoning models burn very different token counts on the same task.

Sources: research.meta.ai / Simon Willison / TechCrunch AI / Hacker News discussion

Claude Code settings files are exposing secrets in public GitHub repos

Ars Technica reports that security researchers at Apiiro found database passwords, API tokens, and root logins inside Claude Code settings files committed to public GitHub repositories. The leak comes from the "Yes, and don't ask again" approval flow, which writes the approved shell command into .claude/settings.local.json, often verbatim and including any embedded credentials. Anthropic's documentation notes the file is added to the developer's global .gitignore, but Ars Technica reports that this is not reflected at the repo level, so the file can be committed when a repo is cloned to another machine or a folder later becomes a repo. Apiiro found over 100,000 uploaded settings files, about 1,100 of them from 63 organizations containing secrets.

Why it matters: Anyone running Claude Code should check whether .claude/settings.local.json is ignored at the repo level and rotate any credentials that were typed into approved commands, since Ars Technica reports secrets can persist in forks even after removal.

Sources: Ars Technica AI

Writeup says consumer Gemini can route Drive files and emails into human review

A GitHub writeup documents that consumer Gemini with Connected Apps enabled can pull Gmail and Drive content into context during ordinary conversation, without an @Drive or @Gmail tag, and that those sessions are subject to human reviewer pipelines. The author points to Google Help Answer 16836988 and an October 16 edit to Google's 2023 Workspace extensions blog post, which previously promised personal content was not seen by human reviewers, and quotes Google's warning not to connect apps containing confidential information. The author says retrieval is probabilistic, ranging from no retrieval to full unredacted ingestion of a document or email. The writeup states the scope is personal Google accounts, and that Workspace Business, Enterprise, and Education accounts are covered by data processing terms that exclude human review and generative training by default.

Why it matters: If you prototype with a personal Google account, assume connected Drive and Gmail content can end up in a reviewed session, and keep confidential data on Workspace accounts or out of Gemini entirely.

Sources: github.com / Hacker News discussion

DeepMind's AlphaProof Code formally verifies the Signal protocol's implementation

Google DeepMind says it used AlphaProof Code, a multi-LLM agentic framework, to mathematically prove key security properties of the actual code behind the Signal protocol. Human experts translated Signal's security goals into formal specifications, then agents powered by Gemini Carbon generated over six million lines of Lean; the agents flagged about one percent of the codebase for human review, with Lean checking the rest. DeepMind says the process took four weeks for work that historically takes expert teams years, and that reintroducing two real, previously fixed bugs produced machine-checked proofs that those versions were insecure. DeepMind says it is open-sourcing the formal verification artifacts.

Why it matters: This pushes formal verification from protocol design down to shipped implementation code, which is where real vulnerabilities tend to hide, and the open-sourced artifacts give teams working on high-assurance code a concrete starting point.

Sources: Google DeepMind

Nvidia board opens independent investigation after Jensen Huang admits affair

Ars Technica reports that Nvidia CEO Jensen Huang acknowledged an extramarital affair in an internal message to employees on Sunday, two days after the Financial Times reported his alleged relationship with a woman US intelligence believes was recruited to spy for China's Ministry of State Security. Nvidia said its board will open an independent investigation "in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort," and Huang said he would fully cooperate. The FT linked the woman to Alice Huang, an executive at Megaspeed, a Singapore-based Nvidia customer. TechCrunch reports that short seller Culper Research flagged the relationship as "uncomfortably close" in a May 13 report, and that Wired reported Huang engaged crisis communications firm Heller Inc. at least a year ago.

Why it matters: Governance and export-control scrutiny at Nvidia is a background risk for anyone whose compute roadmap depends on its chips and its China-adjacent sales channels.

Sources: Ars Technica AI / TechCrunch AI / TechCrunch AI

CNBC: CoreWeave bond yield tops 16% as AI credit stress widens

CNBC reports that the yield on CoreWeave's 2032 bond, which carries a 9.625% coupon, has climbed above 16%, and that its five-year credit default swaps spiked above 1,289 basis points on Thursday before easing to around 1,155 on Friday, per Deutsche Bank. CDS on Nvidia debt is up 76% over the last month, Broadcom's up 49%, and Amazon's up 44%, though CNBC notes those remain low in absolute terms. Credit markets are also weighing the prospect that data center projects are delayed or scrapped under incoming Democratic governors in Ohio, Texas, and Georgia; Goldman Sachs estimates moratoriums could affect 45% of planned US capacity. Goldman also reported that hyperscalers collectively had negative free cash flow in the third quarter.

Why it matters: Higher financing costs for neoclouds tend to show up later as fewer or shorter compute contracts and changed prepayment terms, so teams planning long GPU commitments should watch counterparty risk.

Sources: cnbc.com / Hacker News discussion

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.