Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · November 10, 2026 · afternoon edition

Today's AI news is about the parties you never see in the loop, with Notion customers learning their AI chat transcripts sat at a breached inference vendor, GitHub shipping Claude Fable 5.2 with Anthropic retaining prompts by default, and DeepMind putting a proof checker between AI-written verification and Signal's cryptographic code.

Fireworks AINotion AIAlphaProof CodeGitHub CopilotClaude Fable 5.2agent-securityprivacyfrontier-modelsclaude-codeagent-memory

Trending AI Briefing: Tuesday, November 10, 2026 (afternoon ET)

Every AI feature has more parties in it than the screen shows. Today's news is a tour of them. A Notion feature routed AI chat transcripts to an inference vendor that got breached, GitHub put Claude Fable 5.2 into Copilot with the model provider keeping prompts by default, and Google DeepMind showed the one extra party you actually want in the loop: a proof checker that does not take the model's word for anything.

What's hottest in AI news right now

Notion says some Notion AI chat transcripts were caught up in a security incident at Fireworks AI, the inference company it used behind the scenes. The notice reached the top of Hacker News early on November 10. Notion says Fireworks told it on November 6 that "malicious actors may have accessed potential customer data," and confirmed the dataset on November 8 (a separate Notion security advisory, updated November 9, dates Fireworks' first notice of the incident to the evening of November 2). The exposed material is "unattributable Notion AI chat transcripts used for the Insights feature," a feature that helps customers understand their own AI usage. Notion says that for the vast majority of affected workspaces the transcripts carried no author or workspace labels and no tool-call results, and that fewer than 1% of workspaces may be affected. The catch sits in Notion's own advice: anyone who pasted passwords, API keys or other credentials into Notion AI should rotate them "right away." Notion has disabled Insights, stopped serving open-weight models through Fireworks, and says it disconnected and deleted its Fireworks data as of November 3. Notion notice · HN discussion

Fireworks AI's own bulletin, last updated November 9, describes a wider incident than the Notion slice. Fireworks says a third-party cloud provider hosting its R&D machines had a security incident on October 23, and the attacker used it the next day to reach Fireworks development machines and run "an unauthorized workload." Fireworks now reports impact to account and billing information, and for a limited set of accounts to secrets, datasets and incomplete model weights. It links the attacker to a financially motivated group that sells stolen access to language models through illicit routing gateways, a practice it calls "LLMJacking." Fireworks says it has seen no impact to its Inference, Training SDK and Nexus products, which are zero data retention by default. Its bulletin does not mention Notion or chat transcripts, so the two accounts should be read side by side rather than merged. Fireworks bulletin

Google DeepMind introduced AlphaProof Code on November 10 and used it to formally verify Signal's core cryptographic code. Formal verification means proving, with a machine-checked proof, that code does what its specification says. DeepMind describes AlphaProof Code as a multi-LLM agentic framework that splits that work across agents running on Gemini Carbon, which wrote proofs in the Lean proof language. DeepMind says the job took about four weeks, against the years such projects have historically taken expert teams, and that the agents flagged about 1% of the codebase for human review. When the team reintroduced two previously fixed bugs, the system produced a machine-checked proof that each buggy version was insecure. One number needs care: the post's body says agents wrote over six million lines of Lean, while the page's summary metadata says three million. The verification artifacts are open-sourced. DeepMind post · Artifacts

GitHub made Claude Fable 5.2 generally available in Copilot on November 9, with data retention on by default. It covers Copilot Pro+, Max, Business and Enterprise across VS Code, Visual Studio, the Copilot CLI, the coding agent, JetBrains, Xcode, Eclipse and github.com. Business and Enterprise admins have to switch on a "Claude Fable 5.2 policy," which is off by default. GitHub's changelog says the model "requires data retention by default": Anthropic keeps prompts and outputs to run safety classifiers and does not train on them. Zero data retention is available only to certain approved enterprises through the end of the calendar year, after which it requires Anthropic's Enterprise Frontier Safeguards. Billing runs at provider list pricing under usage-based billing; Claude Code's 2.1.321 changelog entry lists Fable 5.2 at "$8/$40 per Mtok." GitHub changelog · Claude Code changelog

Anthropic published a November 9 research post in which Claude Fable 5.2 designed three working electronic devices. With the hardware startup Diode Computers, Claude produced schematics, circuit board layouts, enclosure CAD, firmware and parts lists for a USB speakerphone, a voice-assistant hub and a handheld virtual pet. Diode built the boards from Claude's files, and Anthropic says the prototypes worked "without any fixes to their circuitry." On PCB Gym, a circuit-board design benchmark, Anthropic reports 69% for Fable 5.2 against 25% for Fable 5. The honest catch is spelled out in the post: none of the devices has been through electrical validation, antenna tuning, thermal testing or certification, and Anthropic has a commercial relationship with Diode. Anthropic research

New tools and features worth actually trying

Notion's "Security incident review" panel. If you own a Notion workspace and were contacted, open the workspace sidebar, then the "?" menu, then "Security incident review" (desktop or web, not mobile) to see the affected transcripts; Enterprise owners get a CSV. Honest tradeoff: Notion's notice does not define the exposure window, so the safe move is to rotate any credential anyone ever pasted into Notion AI, not just the ones in the file.

AlphaProof Code's open verification artifacts. If you write or audit security-critical code, the published Lean proofs are the most concrete look yet at what AI-driven formal verification produces at scale. Honest tradeoff: this is a research release, not a product; there is no tool to point at your own repo, and reading Lean proofs takes real background.

Claude Code 2.1.322. The current latest on npm adds fast mode on Google Vertex AI in the global, us and eu regions, falling back to standard speed if Vertex refuses it. Update with npm install -g @anthropic-ai/claude-code@latest. Honest tradeoff: npm's stable tag still points at 2.1.311, so teams that pin to stable should not expect this yet.

Codex CLI 0.167.0's codex agents --no-daemon. Released November 6, it adds a command center for running agents without a background daemon, with unread markers and backend switching in the session picker. Honest tradeoff: it is four days old and outside today's window, and OpenAI's changelog lives at a redirected URL that renders inconsistently.

Trending AI repos on GitHub today

Trendshift's board was read at about 15:10 ET; it showed no rank numbers or scores, so the numbers below are positions in its page order and its figures are momentum signals, not verified totals. Star counts came back inconsistent between shields.io and GitHub's pages this run, so they are given as rough ranges.

  • ranxianglei/billion-context (#23): a proxy that sits between a coding agent and its model API and lets the model fold old conversation into summaries it can reopen later, aimed at sessions that run for weeks. Why now: v0.1.213 shipped November 7 and the repo is climbing fast. MIT plus an extra attribution term for user-facing products; star counts disagree between shields.io (about 6.2k) and the repo page (about 4.2k). Caveat: the README calls the project "Early" and tested against mocks while also citing a self-authored study of 174,327 real model calls, and its benchmarks are the project's own.
  • amosblomqvist/learnharness (#8): a desktop app where AI agents assess what you know, map a path to a goal, then teach and quiz you, keeping that knowledge across sessions. Why now: v0.3.7 landed October 25 and it sits near the top of the board. AGPL-3.0, several thousand stars (shields.io and the repo page disagree). Caveat: you bring your own model key or a Claude Code install, and the README hints at future paid managed keys or hosting.
  • h4ckf0r0day/obscura (#9): a headless browser engine in Rust, pitched as a drop-in for headless Chrome under Puppeteer and Playwright for scraping and agents. Apache-2.0, roughly 30k stars, latest release v0.2.5 (October 17). Caveat: its speed figures are self-reported, two README tables disagree with each other, and the README carries sponsor promo codes for a proxy service.
  • anthropics/skills (#11): Anthropic's public repository of example Agent Skills, folders with a SKILL.md that Claude loads for specialized tasks. Well over 180k stars, no releases. Caveat: there is no root LICENSE file, so licensing lives skill by skill; check the folder you copy, and note the README frames the skills as demonstration and educational material.
  • shinyobjectz/raea (#14): "Render Anything Engine," a canary-stage engine that renders video, motion graphics and 3D written in React and TypeScript on your own GPU. Star figures conflict between sources; no releases. Caveat: it uses the Functional Source License, which bans competing commercial use and only converts to Apache-2.0 two years after each release, and the repo had two commits when checked.
  • google/elephant-goldfish-model (#15): a plugin for coding agents in which a context-rich agent drafts a spec and fresh-context subagents review it before a human approves. Apache-2.0, star figures conflict between sources (shields.io about 17k, the repo page about 13k), no releases. Caveat: the README says it is "not an officially supported Google product" despite the org name.
  • kingcharleslzy-ai/dsh-purge (#17): an unofficial plugin for the DeepSeek Harness desktop app that folds its repeated confirmation prompts into one setting so long agent runs stop less often. MIT, several thousand stars (sources disagree), v1.6.88 released today. Caveat: the README nicknames it "armor-breaking" and tags it "jailbreak," it patches the host app, and it contradicts itself on patch counts and current version.

What actually matters from today's signal

Track the second party. Notion's users chose Notion, not Fireworks, yet a Notion analytics feature put their chat transcripts on Fireworks infrastructure, and Fireworks' attacker came in through yet another company's cloud. The highest-signal areas for builders this week: the sub-processor list behind every AI feature you ship or buy; retention defaults on frontier models, which GitHub's Fable 5.2 page shows now ship on by default; anything that sits in your model API path, a category that includes popular compression proxies like billion-context; and verification you can check yourself, which AlphaProof Code's proofs make concrete.

The counter-signal is that the story everyone is reading as "Notion got hacked" is really about secondary uses of AI data. Notion's notice ties the exposed transcripts to Insights, the analytics feature that summarizes how people use Notion AI, rather than to the assistant's answers. Every product team building usage insights on top of chat logs is creating the same copy. And the feel-good half of the news deserves the same skepticism: Anthropic's devices are unvalidated prototypes built with a commercial partner, and DeepMind's verification numbers disagree with themselves on a single page. Read the caveats first. They are the actual news.

Source access notes: Vendor pages, Hacker News (Algolia stories since 08:11 UTC November 9 plus the front page), Hugging Face papers, Trendshift and repo files were read between about 15:10 and 15:50 ET. openai.com/news failed to load; the Codex changelog redirected to learn.chatgpt.com and was read there. blog.google's AI index returned a stale page, and Google's "Gemini 4 Carbon is now available" post (on Hacker News at 16:07 UTC November 10) returned an error twice, so whether access widened beyond the cyber defenders named in the October 28 launch post could not be confirmed and the item was left out. DeepMind's blog index shows month-only dates; AlphaProof Code's date comes from its own page. Notion's notice carries no publish date; dates come from its body. Notion's sub-processor list failed to load, so whether Fireworks was listed is unconfirmed. npm's time map was truncated, so 2.1.322's publish time is not stated. Trendshift showed no ranks or scores. All releases.atom feeds were blocked by robots rules; the repo pass used /releases/latest, which omits the year (2026 assumed from context), and raea's release page showed no releases. Adversarial pass (Sonnet 5.5) caught: a Fable 5.2 price split the Claude Code changelog does not state, an anthropics/skills "source-available" claim the raw README contradicts, an overstated description of dsh-purge (it consolidates confirmations rather than removing them), star counts that disagree across sources (now given as ranges), raea's release status, and an interpretive Notion claim stronger than the notice (softened). The Codex fact check on the Notion article (Step 11) surfaced Notion's separate security advisory, which limits "unattributable" to the vast majority of affected workspaces and dates Fireworks' first notice to November 2; both points were folded into this briefing and the X-article before publishing.