Anthropic: Claude formally verifies 123 open-source libraries, writing ~180M lines of Rocq proofs
Scales formal verification from niche expertise to 100+ libraries; fresh flagship research.
AI NEWS
Past AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
This day
October 23, 2026
17 items for 2026-10-23, ranked by combined engagement. Unverified items are marked.
Top 3
Changes daily UX friction for paid Claude users; strong engagement.
Moves an agent interoperability standard into neutral multi-lab governance.
All
Anthropic: Claude formally verifies 123 open-source libraries, writing ~180M lines of Rocq proofs
Scales formal verification from niche expertise to 100+ libraries; fresh flagship research.
CollapseAnthropic: Claude formally verifies 123 open-source libraries, writing ~180M lines of Rocq proofs
Anthropic’s Frontier Red Team, with Theorem, used Claude over a few weeks to prove security properties at the RISC-V assembly level for 239 entry points across 123 real open-source libraries—showing no crash from memory-safety errors or failed assertions for any input under 64 MB under stated assumptions (single-threaded static binary, fixed memory budget, Linux on RISC-V). The effort produced about 180 million lines of Rocq (typical proof ~600k lines on binaries with median ~18k RISC-V instructions). While proving, Claude also found bugs that years of fuzzing missed. Anthropic is publishing a subset of proofs and RISC-V/OS semantics; Theorem offers the same pipeline for other OSS and is shipping a Claude skill.
Claude web, desktop and mobile switch to Auto Mode by default next week; classifier blocks risky actions
Changes daily UX friction for paid Claude users; strong engagement.
CollapseClaude web, desktop and mobile switch to Auto Mode by default next week; classifier blocks risky actions
Anthropic says that starting next week, Claude sessions on web, desktop and mobile for Pro, Max and Team plans will default to auto mode: instead of approving every step, an automated check blocks risky actions, and users can switch back to manual anytime. It cites Claude Code testing where auto mode caught about 89% of dangerous commands versus about 14% under manual approval. Free users are unaffected. The change extends the August Claude Code default to full-app sessions.
Anthropic donates Agent Skills to the Agentic AI Foundation; 70+ platforms already integrate the standard
Moves an agent interoperability standard into neutral multi-lab governance.
CollapseAnthropic donates Agent Skills to the Agentic AI Foundation; 70+ platforms already integrate the standard
ClaudeDevs announced donating the Agent Skills standard to the Agentic AI Foundation—the same home as MCP. Since December 2025, 70+ agent platforms have integrated it, including ChatGPT, Cursor, GitHub Copilot and VS Code. Existing skills and licenses stay the same; anyone can use or contribute without joining AAIF. A Technical Steering Committee with engineers from Anthropic, Earendil (Pi), Google, Microsoft, OpenAI and OpenCode will govern it, prioritizing versioned releases, a formal proposal process and security guidance.
Richard Socher launches Flint, Recursive’s first local coding model: 125+ tok/s on MacBook; open weights
High-reach local coding launch with open weights and concrete speed numbers.
CollapseRichard Socher launches Flint, Recursive’s first local coding model: 125+ tok/s on MacBook; open weights
Recursive SI launched Flint, its first model: a fast local coding model built from scratch in under four months with its Eureka Machine automated research system. Claims include best-in-class fast local coding and up to ~2× speed versus comparable models—125+ tokens/sec on a MacBook Pro and 150+ on RTX Spark. Open weights are on Hugging Face for local runtimes; free trials are available on OpenCode and pi.dev. Pitch: private, unmetered intelligence on your own machine.
1Password for Agents: agents request credentials; on approval 1Password auto-fills so agents never see passwords
Fills a critical agent-era credential gap with multi-vendor partnerships.
Collapse1Password for Agents: agents request credentials; on approval 1Password auto-fills so agents never see passwords
1Password launched its Agents platform with partners including Muse, xAI Grok Bot, Claude, OpenAI, Perplexity, Browserbase and others. When an agent hits a sign-in form it requests a credential from your vault; if you approve, 1Password auto-fills the destination so the agent never sees or holds the password. Free and optional for personal/family accounts; agents cannot touch the vault unless you opt into a connector. Aimed at users who want personal AI agents without giving up password control.
Claude models on Google Cloud (Anthropic-managed) reach general availability with IAM, billing and commitment drawdown
Another major cloud distribution path for Claude enterprise buyers.
CollapseClaude models on Google Cloud (Anthropic-managed) reach general availability with IAM, billing and commitment drawdown
Anthropic said Claude Models on Google Cloud (Anthropic managed) is generally available: the full Claude Platform feature set on Google Cloud—including Managed Agents, code execution and Opus 5.5 fast mode—with same-day feature parity. It uses Google Cloud IAM credentials, draws down GCP commitments, and offers US and Global endpoints. Enterprises can adopt Claude inside existing GCP billing and identity setups.
OpenAI models coming to Google Cloud Model Garden; buyable with GCP commitments; Microsoft remains primary partner
Puts major closed models on one GCP shelf; big enterprise-choice story.
CollapseOpenAI models coming to Google Cloud Model Garden; buyable with GCP commitments; Microsoft remains primary partner
OpenAI and Google Cloud will bring OpenAI frontier models and Codex agentic workflows into Model Garden so enterprises can use them where they already build, with GCP access controls and data governance. Purchases go through Google Cloud Marketplace and can draw down GCP spend commitments. Models will also appear in Google Antigravity and the Gemini Agent. General availability is planned in the coming weeks; Microsoft remains OpenAI’s primary cloud partner.
Claude Security adds continuous re-scans on code changes; Enterprise public beta with up to $25K/mo credits through 2026
Turns security scanning into continuous, change-aware coverage.
CollapseClaude Security adds continuous re-scans on code changes; Enterprise public beta with up to $25K/mo credits through 2026
Claude Security now connects once to GitHub, GitLab and other Git providers, runs a deep first scan, then re-scans as changes land so findings track current code. When a fix ships, Claude verifies and marks it fixed even if fixed outside Claude; regressions reopen. Each project gets an editable threat model. Findings reach tools via a new API and MCP connector. Public beta on Claude Enterprise; eligible orgs get up to $25K monthly usage credits for two months through Dec 31, 2026 (terms apply).
OpenAI GPT-Rosalind: ~2,000 agents analyze a mouse whole-brain CRISPR atlas in 32 minutes; cholesterol-transport drug lead
Concrete multi-agent science result with a public atlas and testable leads.
CollapseOpenAI GPT-Rosalind: ~2,000 agents analyze a mouse whole-brain CRISPR atlas in 32 minutes; cholesterol-transport drug lead
OpenAI, PerturbAI and Scripps Research ran thousands of GPT-Rosalind agents in parallel on a whole-brain CRISPR atlas of millions of mouse brain cells, finishing in about 32 minutes work humans estimated would take months to a year, and published a reasoning atlas. Agents linked DCTN1 (Perry syndrome) disruption to higher cholesterol-synthesis gene expression, suggesting impaired cholesterol transport as a testable neurodegeneration lead, and flagged cerebellar differences between TSC2 and TSC1 disruptions relevant to tuberous sclerosis. Blinded experts preferred distributed-agent reports about 75% of the time.
Canva launches Craft, its design-focused frontier image model; claims #2 on DesignArena’s global leaderboard
Major design platform ships its own frontier image model; huge reshare volume.
CollapseCanva launches Craft, its design-focused frontier image model; claims #2 on DesignArena’s global leaderboard
Canva introduced Craft, its first frontier image-generation model built specifically for design, saying extraordinary image generation is now available to everyone. With Craft, Canva claims it is the world’s second most advanced AI lab for image generation on DesignArena’s global leaderboard. The launch includes a demo video and pitches generation inside design workflows rather than as a generic text-to-image tool.
DeepMind releases ForTy: a global 10 m forest-typology map of natural, planted and farmed forests
Moves beyond tree cover to forest type—key for regulation and carbon accounting.
CollapseDeepMind releases ForTy: a global 10 m forest-typology map of natural, planted and farmed forests
Google DeepMind with Google Research, WRI and IIASA released Forest Typology (ForTy) 2020: a global map that sorts every 10×10 m patch into six FAO-based types—primary, naturally regenerating, planted, plantation, tree-crop/agroforestry and other land—plus an early 2017–2025 time series showing where natural forest was replaced by plantations or farms. Built on AlphaEarth Foundations reading multi-year satellite imagery across ~1.7M sites, ForTy 2020 and the series are free on Google Earth Engine for deforestation monitoring and deforestation-free supply chains (including EUDR baselines). Paper in Scientific Data.
[Rumor] Bloomberg: DeepSeek raises ~¥100B (~$15B), led by Tencent and CATL; company has not confirmedUnconfirmed
Would be a landmark China AI raise if true; still unconfirmed by the company.
Collapse[Rumor] Bloomberg: DeepSeek raises ~¥100B (~$15B), led by Tencent and CATL; company has not confirmed
Per Bloomberg’s Oct 21 exclusive as summarized by @dotey: sources say DeepSeek raised about ¥100B (~$15B), up from talks that started near ¥50B then ¥80B; largest backers reportedly include Tencent and CATL. It would be DeepSeek’s second external round in six months after ~¥51B in June. Caijing reported ~¥500B pre-money valuation, CATL planning ~¥100B more, new names like Geely and SMIC-linked funds, and a possible STAR Market path via CITIC. DeepSeek, Tencent and CATL have not publicly confirmed—marked as rumor.
OpenAI hosted METR and Redwood for a three-week embedded eval of internal agent monitoring; brief now public
Third-party on-site eval of internal agent monitoring—governance signal.
CollapseOpenAI hosted METR and Redwood for a three-week embedded eval of internal agent monitoring; brief now public
OpenAI said it invited METR and Redwood Research this summer to pilot an embedded evaluation of its internal monitoring system with deep infrastructure access for three weeks. METR said two employees worked on-site assessing systems OpenAI uses to monitor misalignment from internal agents; a brief write-up is now public. OpenAI says the findings helped strengthen safeguards and will shape future collaborations with METR and other third-party evaluators.
Gemini 4 Argon builds an interactive 3D human-anatomy visualization from three open-source datasets
Clear multimodal demo of frontier Gemini on scientific visualization.
CollapseGemini 4 Argon builds an interactive 3D human-anatomy visualization from three open-source datasets
Google DeepMind showed Gemini 4 Argon combining three open-source datasets into an interactive 3D visualization of human anatomy—toggling layers in motion, isolating muscles, or zooming inside a beating heart. A capability demo after the Gemini 4 Argon launch, highlighting multimodal and 3D scene construction.
Artificial Analysis launches CritPt v2; GPT-6 Astra (max) scores 83%; coming to the Intelligence Index
Refreshed research-physics benchmark with clear frontier model rankings.
CollapseArtificial Analysis launches CritPt v2; GPT-6 Astra (max) scores 83%; coming to the Intelligence Index
Artificial Analysis launched CritPt v2: all 70 research-level physics challenges were revised, and models now solve them agentically with a code-execution tool in an offline sandbox. Leaders: GPT-6 Astra (max) 83%, GPT-6.1 Sol (max) 82%, Claude Opus 5.5 (max with fallback) 78%, GPT-5.6 Sol (max) 75%; GPT-5 (high) from Aug 2025 scores 29%. v2 scores are higher than v1 and not directly comparable; AA says most of the jump is from revised challenges, with tool access a smaller gain. CritPt changes will enter the Intelligence Index update in late October.
Ant Group Ming-3.0-tts: 0.5B MIT-licensed TTS ranks #1 open-weights on AA Controlled Voice Arena
Tiny open TTS punches into top arenas—strong CN open-weights signal.
CollapseAnt Group Ming-3.0-tts: 0.5B MIT-licensed TTS ranks #1 open-weights on AA Controlled Voice Arena
Artificial Analysis says Ant Group’s Ming-3.0-tts (dense 0.5B; plus 7.9B MoE Ming-3.0-tts-tiny) is #1 open-weights on the Controlled Voice TTS Arena (Elo 1,069) and ties Fish Audio Drama 3 Preview at 93.2% on Pronunciation Robustness, ahead of Eleven v4 at 91.7%. It is the only sub-1B model in the top 3 of any AA TTS board; every other open-weights model in Controlled Voice top 10 is 3.5B+. Weights are MIT-licensed on Hugging Face.
Logan Kilpatrick: Gemini 4 Argon processed ~25 trillion tokens in its first 72 hours
Public throughput number quantifying launch demand for Gemini 4 Argon.
CollapseLogan Kilpatrick: Gemini 4 Argon processed ~25 trillion tokens in its first 72 hours
Google AI Studio lead Logan Kilpatrick said Gemini 4 Argon processed about 25 trillion tokens in the first 72 hours after broad launch, calling demand for intelligence incredible and still under-appreciated. The figure was widely reshared, including by Demis Hassabis, as a public cold-start demand signal.