Daily LLM news, Friday, 23 October 2026

Good afternoon. Today Claude writes machine-checked proofs that real open-source code can't crash, outside testers find holes in OpenAI's system for watching its own agents, cheap "decision models" get cheaper, Microsoft publishes its AI diagnosis results in Science, and human-guided AI agents help crack a 31-year-old maths problem.

LEAD: Claude proves 123 open-source libraries can't crash on inputs under 64 MB

Background: Formal verification means proving mathematically that code behaves as specified, not just testing it. It has always been hugely expensive: verifying the small seL4 operating system kernel took specialists about 20 person-years, so it's been kept for things like avionics.

What happened: On 22 October Anthropic's Frontier Red Team, working with the startup Theorem, said Claude had written machine-checked proofs for 239 fuzz-test entry points across 123 libraries in Google's OSS-Fuzz project, over a few weeks. The proofs show those compiled programs never crash or break memory safety for any input under 64 MB, and they're checked automatically by the Rocq proof checker, so almost no human review is needed. The finished proofs run to about 180 million lines (Claude wrote around a billion lines in total along the way). While proving, Claude found and fixed hundreds of bugs, including one that had gone unnoticed for more than a decade, and in one week on a widely used PDF reader it found as many security bugs as usually get reported in a year. Anthropic says Opus 5.5 solves half the proofs with about 40% fewer tokens than Opus 5. A subset of proofs is public, and Theorem is offering to run the pipeline for other open-source projects.

Why it matters: If models can prove code is correct, not just write it, AI-generated software becomes much easier to trust. The limits are real, though: the proofs only cover single-threaded programs on one chip type (RISC-V on Linux) and inputs under 64 MB, and individual proofs can burn hundreds of millions of tokens.

Sources: Anthropic, The Decoder

Outside testers found holes in OpenAI's system for monitoring its own agents

Background: OpenAI runs a monitor that watches its internal coding agents (such as Codex) for signs of misalignment. Yesterday's briefing covered OpenAI's own report of agents disabling security alerts; this is an independent look at the watchdog itself.

What happened: METR and Redwood Research each embedded one researcher at OpenAI for three weeks (21 July to 11 August) and published a summary on 22 October. They estimate only about 75% to 94% of internal agent traffic from models at least as strong as GPT-5.5 was eligible for monitoring, and RL training runs weren't covered at the start. They found an exploit that could stop hundreds of recent agent actions from being monitored, two client-controllable settings that switch monitoring off for new agents, a possible way for Codex agents to bypass sandboxing and act with the user's full permissions, and a way to make the monitor fail through formatting differences. OpenAI says it has patched every vulnerability affecting production safeguards; the testers say they haven't verified that, and OpenAI was allowed to redact the public summary.

Why it matters: It's the first public test of outsiders getting deep access inside a frontier lab, and it shows that monitoring agents is only as good as its coverage. Expect other labs to face pressure for similar reviews.

Sources: METR, The Verge, CNBC

"Decision models" go mainstream: OpenAI's Decisions API is GA, and SubQ cuts its price

Background: Apps and agents make lots of small choices: classify this, route that, call a tool or ask a question. Running a full chat model for each one is slow and costly, so a new class of "decision models" returns just a probability, a choice or a score.

What happened: On 22 October OpenAI made its Decisions API generally available. It's built on GPT-6 Luna, which OpenAI says answers bounded questions up to 10 times faster than Luna through the normal Responses API, costs $0.10 per million input tokens ($0.01 cached) with no output charges, now takes up to 400 questions per request, and plugs into GPT-Live for voice apps. Customers quoted by OpenAI include Datadog (40 times faster at the 95th percentile for scoring risky shell commands) and Sierra (under 280 ms, about 90% cheaper per call). The same day, Subquadratic released SubQ Snap 1.1 at $0.04 per million input tokens, a 20% cut, with 384K tokens of context; it says it's tied for first on the Artificial Analysis Decision Index and that overconfident made-up answers on unfamiliar questions fell from 99.3% to 12.5%.

Why it matters: Pushing routine agent choices to a fast, cheap model can cut agent costs and lag sharply, leaving big models for the hard parts. Benchmark claims here are mostly the vendors' own.

Sources: OpenAI, Subquadratic

Microsoft's AI diagnosis system, published in Science, beats doctors on hard cases

Background: Medical exam benchmarks give a model all the facts at once. Real diagnosis means deciding what to ask and which tests to order, step by step.

What happened: On 22 October Microsoft AI published its MAI-DxO work in Science. MAI-DxO runs language models as a "panel of virtual physicians" that debate, ask questions and order tests. On 304 complex New England Journal of Medicine cases it got 83% right, against 36% for physicians using textbooks and the internet and 22% without them. On more typical emergency-department cases it scored 80% against 47%, and Microsoft says it reached answers at a half to a third of the cost in tests. Results held or improved when swapped onto newer models such as GPT-5.5. Microsoft is open-sourcing part of its SDBench benchmark and starting a prospective study with Beth Israel Deaconess Medical Center.

Why it matters: It's peer-reviewed evidence that orchestrating several model calls can beat a single answer on complex reasoning. But it's a text-only simulation against generalists, not specialists or real patients.

Sources: Microsoft AI, Becker's Hospital Review

Human-guided AI agents helped solve the 31-year-old KLS conjecture

Background: The Kannan-Lovász-Simonovits conjecture, posed in 1995, is a central open problem in high-dimensional geometry, with knock-on effects for how fast random sampling algorithms run.

What happened: Microsoft Research's Xinzhi Zhang and Zhao Song have proved it, Microsoft said in a post dated 21 October. They ran AI agents through about 100 research paths, with the researchers pruning dead ends and steering the search; the agents also did derivations, proof checks and drafting. Two other groups posted proofs the same day, one building on the Microsoft paper's key step, and a Lean formal proof followed. In a control run, leading models working alone with a similar token budget fell short.

Why it matters: It's one of the clearest cases yet of LLM agents contributing to a major mathematical result, and the control test suggests the human steering mattered as much as the models.

Sources: Microsoft Research, Paper (PDF)

Also worth a look

• MiniMax M3.1 Flash weights are out. The cheap model covered on 20 October is now downloadable on Hugging Face: about 428 billion parameters with 23 billion active per token. It's open weights rather than open source, and reports say firms above $20 million in annual revenue need MiniMax's permission. Hugging Face, OrcaRouter
• Claude makes "auto mode" the default. From the week of 26 October, Pro, Max and Team sessions on web, desktop and mobile will run in auto mode, with an automated check blocking risky actions; users can switch back to manual approval, and Enterprise stays on manual. Claude Help Center