Mistral's 1T-param Le Chonk 🐈, GitHub rebuilds Git for agents 🏗️, Meta's agent login standard 🪪

The public preview is available through an API on Mistral Studio, at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the model scores 82% on a test of reproducing and patching real open-source vulnerabilities, the highest of any model, while closed models including Claude Opus 5.5 score near zero because they refuse the task. Until the weights arrive, Mistral is red-teaming it with cybersecurity leaders, vetted partners and state authorities.

Microsoft has cut projected internal Claude spending, once at least $1 billion, by more than a third after asking staff to use its own tools, sources said. Meta’s Claude Code users fell from about 60,000 to about 30,000, according to a source and internal data, partly through layoffs and more through in-house coding tools. Microsoft’s total payments to Anthropic have stayed roughly flat as Claude-powered Copilot features for customers grew, one source said.

South Korean officials said the attacks, first detected last week, hit at least seven financial firms and took the personal data of 68,000 people. Investigators detected traces of Artex, an open-source security agent built by a Chinese engineer, which draws on models such as Opus, GPT or DeepSeek to locate vulnerabilities and plan attack routes with minimal human input. No culprits have been identified, and officials did not say which models were used.

OpenAI posted 372 results, saying each resolves or makes substantial progress on a major open question in mathematics or theoretical computer science. A spokesperson told Scientific American an internal model, not yet public, produced almost every one from a single prompt to a single agent, a claim MIT’s Andrew Sutherland says to treat as unverified. Many are verified in Lean, a language that validates a proof’s logic, but OpenAI is not disclosing its prompts.

The endpoint takes text, images or both and answers three kinds of question: the probability that a condition is true, one option from a fixed set, or a score against a rubric. Only gpt-6-luna is available, at $0.10 per million input tokens with no output-token charge, and OpenAI expects general availability in the coming weeks. Its guide suggests using the answers to classify content, route requests and prioritise work.

Personal Agent Protocol is being built with partners including Genesys, Shopify, Stripe and Walmart, with a v0.1 specification planned for later this month. An agent discovers on a company’s website what it offers, then starts a session built on OAuth, an established standard for authorising access, and the customer decides whether it gets read-only or write access. The company chooses whether the agent works through its web pages, its APIs or its own agent.

The Cyber Verification Program merges with Project Glasswing and covers Opus 5.5, Sonnet 5.5 and Mythos 5.1. Defense Access is for defensive work such as incident response and reverse-engineering malware, Red Team Access adds authorised penetration testing for organisations only, and Specialized Access is reserved for a limited set reviewed with the US government. In Anthropic’s own test, the Defense tier still blocked 46 of 50 offensive trials, and enrolled organisations must accept data retention.

GitHub's current storage, Spokes, keeps full copies of each repository on several fileservers, and every copy takes part in every write, so adding replicas to absorb reads makes pushes slower. The new design holds the authoritative copy in Azure Blob Storage, serves reads from lightweight caching workers and limits coordination to the reference update in each push. In internal benchmarks it delivered up to 35 times higher write throughput, GitHub says.

Before mid-training its Kolibri model, Aleph Alpha scanned 8.3 billion documents against every benchmark it evaluates on and removed 1.5 million that quote a benchmark item nearly verbatim. At the pool's natural leak rate, about 0.04% of tokens, a small stand-in model showed no score gap. Adding every removed document back exactly once, which Aleph Alpha calls an upper bound for most sources, also raised MMLU by 28 points, including on questions that never leaked.

Ronacher, who works on the Pi coding agent, describes Codemode as code the model writes to combine tool calls inside the harness, the program that supplies tools to the model, so intermediate results need not pass through the model's context. Pi runs it with no network, file system or timers. He says MCP servers are not really built for it yet: many do not return structured JSON, and some change their output with result size.

Writing up his LDX3 keynote, the Pragmatic Engineer author cites unpublished GitHub data with a ninefold rise in agent-authored pull requests in eight months, and Linear data in which agent-created issues have outnumbered human-created ones since July. He says human code review has become theatre, because no engineer can keep up with 5-10 times more code. Non-engineers, he adds, are still not putting code into production.

A METR researcher, helped by an AI agent, took about 10 minutes to uncover a flaw in the transcript viewer of Inspect, a widely used safety evaluation framework, that could have let an agent under test change what a human reviewer sees. METR says it has not observed agents exploiting it, the underlying transcripts were not modified, and Inspect's maintainers patched it within one day. Its advice is to treat agent outputs as untrusted input.

Ross, who concedes the title is a bit hyperbolic, writes that when generating a pull request takes ten seconds and reviewing it properly takes thirty minutes, the work shifts onto volunteer maintainers. His evidence is a run of closed doors. Sindre Sorhus and Hono disabled external pull requests, Ladybird stopped accepting public ones, and curl ended its bug bounty after real reports fell below 5%.

Krishnan had GPT, Claude, Grok, Gemini and Composer write rulebooks for households competing over food in a simulated town of 16 houses, then changed the rules one at a time. Removing the score brought nobody's values back, but letting models send a public message before writing their rulebooks cut theft from about 90 a game to 1.4. He concludes that agents respond to incentives and enforcement, so governing them means designing institutions.

Google says the open model loads only the encoders an index needs, from 270M parameters for text and code to 740M for everything, and images or audio can be added later without re-computing embeddings. Truncating vectors from 768 dimensions to 128 cuts storage sixfold, with text and code keeping around 90% of quality and image, video and speech retrieval around 75%. Google reports a 14% higher score than EmbeddingGemma 1 on the MTEB code benchmark.

The ArtCraft team's open-source image editor is a clean-room reimplementation, written from public specs and observed behaviour, with layers, masks, type and real PSD files. Every menu item, tool and dialog runs from one command registry, so a headless CLI can open, edit and save a file or apply one action list to a folder of images. It is in early alpha and, the README says, not yet a Photoshop replacement for daily professional work.

Setup registers the tool with agents including Claude Code, Codex and Cursor, which can inspect an app without its source code, explain how a feature works and present the evidence. Analysis runs locally, and deep native binary analysis requires Hopper, Ghidra or IDA Pro, while static JavaScript inspection needs only Node and npm. The 6 October release adds static Android APK analysis through headless JADX, and REA does not claim to recover original source code.

The README asks two questions: how far AI agents can go at hardware design, and whether they can build the chip that runs their own inference. On an Inspur YPCB-00338 card, LFM2.5-230M with int8 weights decodes at 59.0 tokens a second on the device, 52.3 including the host, and Phi-4-mini at 3.99, producing the same tokens as the simulator bit for bit. Everything except the card runs on a laptop, including chat on the simulator.