AI Daily · Nov 4, 2026 · Wednesday
BAD INFLUENCE
Anthropic introduces FlashProve for verifying whether AI responses came from a given model
Anthropic has introduced FlashProve, a tool that lets developers verify whether a given response came from a given AI model, building on established cryptographic techniques. The company says it is substantially more efficient than prior open-source methods, and that proofs could make chat hijacking harder and let safety evaluators confirm they are testing the model served to customers. The code is open-sourced on GitHub.
Industry news
The DecoderNews1 more sourceNvidia asks cloud customers to limit business with Chinese companies
Nvidia has told some cloud customers to take on less business from Chinese companies, according to The Information, citing three people familiar with the matter. The guidance is not mandatory, applies to future contracts, and asks customers to disclose the identity of end users accessing the chips.
The Verge · AINews3 more sourcesGoogle will turn off Search for K-12 schools that reject its AI features
Google says it will disable Search for K-12 student accounts unless administrators explicitly enable Gemini and related AI features. Schools cannot turn off the AI features while keeping Search, according to Google education VP Chris Phillips. The change follows skepticism and lawmaker scrutiny over AI in schools, and Google will test a Guided Learning feature in some US schools this month.
Together AI@togethercomputeOfficial1 more sourceRevolut moves about 2,000 engineers to self-hosted GLM for coding
Revolut has made self-hosted GLM its default for coding for about 2,000 engineers, according to David Abitbol, who leads AI operations there. He says the setup runs on 56 Nvidia B200s at a blended 30 cents per million tokens, roughly one-tenth of what Revolut paid Anthropic and OpenAI, and that engineering lead time improved 70%.
fal@falOfficial1 more sourcefal raises $1.65B across Series E and F rounds led by Sequoia and Bessemer
fal announces it raised a combined $1.65B across its Series E and Series F rounds. Sequoia led the Series E, Bessemer Venture Partners led the Series F, and Nvidia continued its investment. The company says the funding will help it push the boundaries of what is possible next year.
OpenAI@OpenAIOfficial3 more sourcesOpenAI plans AI-driven security sprint with Cathedral Systems and raises bug bounties
OpenAI says it is planning its next internal security sprint and partnering with Cathedral Systems, whose researchers will use OpenAI's models to challenge its defenses. OpenAI also more than quadrupled bug bounty rewards for critical findings, with exceptional findings eligible for up to $172,000 including bonuses.
Papers & research
Mistral AI@MistralAIOfficialFollow-upMistral AI releases Mistral Large 4 as open weights with a technical report
Mistral AI says its Mistral Large 4 model is now open-weight and more efficient and compact than its Preview version released three weeks earlier. The company published a technical report on Hugging Face and arXiv describing how the model was built.
Hugging Face proposes Model Integration Contracts to catch model bugs on PRs
Hugging Face proposes Model Integration Contracts, a standard for packaging small test fixtures, usage paths and recorded baselines for Transformers models. The post says a UMT5 bug that let the decoder attend to future tokens went unnoticed on main for almost two months, and that a shrunk UMT5 fixture on CPU flagged it on every replayed commit.
Google Cloud · AI & Machine LearningOfficialGoogle Cloud releases Physical-Engineering's Last Exam, a benchmark for engineering agents
Google Cloud, with Mercor, announces Physical-Engineering's Last Exam (PLE), a benchmark of 85 long-horizon tasks across mechanical, electrical, civil and structural engineering. It releases 29 public tasks, grading tools and an open leaderboard at ple-bench.com. In tests of 14 frontier models, all scored below 50%, with the best reaching a mean of 48.3%.
OpenAI finds agent safety evals and production behavior diverge, driven by objective awareness
OpenAI's alignment team compared agent success on Respecting Warnings evaluations with matched Codex production decisions across six models. Evaluation success was lower than production success for every model, with gaps of 10.2 to 37.4 percentage points, and rank order stayed positively correlated (Spearman's ρ = 0.771). The gap tracked models' verbalized specific evaluation objectives more closely than general awareness that a test was underway, and a system cue naming the wrong objective cut success from 63.4% to 17.6% on GPT-6.1 Sol.
Ai2 (Allen Institute for AI)Official2 more sourcesStateBench tests how language models track state changes over long documents
Ai2 introduces StateBench, a benchmark that tests how well language models track entity state changes in long documents across four difficulty levels. The team reports that recurrent and hybrid models generalize much farther than transformers when trained directly on these tasks. It also finds that frontier models, including Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna, still break down on harder state tracking as inputs grow longer.
Tips & takes
Cognition Blog (Devin, Windsurf)OfficialCognition outlines security steps after the Binance hack and Cursor extension compromise
Cognition says recent incidents, including a $6.34 billion Binance theft and a malicious Claude Code Cursor extension, relied on known attack paths. It recommends immediate steps such as phishing-resistant MFA, removing production secrets from developer machines, and setting a minimum package age of 7 days, with further steps due within 30 days.