AI News Report, October 23: CLAUDE WROTE MATH PROOFS FOR 123 OPEN-SOURCE LIBRARIES AND FOUND HUNDREDS OF BUGS

Anthropic and the startup Theorem used Claude to write machine-checked math proofs that 239 test entry points across 123 open-source libraries cannot crash on any input under 64 MB, surfacing hundreds of bugs that years of fuzz testing missed. The work is the lead item in an October 23 AI news roundup that also covers Google's credential-swapping Agent Gateway preview, a Recursive model with 2.5B active parameters and an 89.7% SWE-bench Pro v2 score on Recursive's own test, and METR's three-week test finding an exploit that could hide hundreds of agent actions and two headers that disabled monitoring, which OpenAI says is patched.

Agents hold no credentials inside the sandbox; Google swaps them in at its Agent Gateway instead. Preview opens in the coming months, and the sign-up form is open now. OpenAI says answers that took one to three seconds now come back in about 100 to 180 milliseconds, and voice and image input are new. Try it for quick pick-one-option steps where a full chat call is too slow. Only 2.5B parameters work per token, and Recursive says it tops 125 tokens a second on an M4 Pro MacBook Pro. It is Apache 2.0, and its 89.7% SWE-bench Pro v2 score is Recursive's own test. Every file, system call and network request is checked against your policy, and agents never see real credentials. It is Apache 2.0 and runs on Linux or an Apple Silicon Mac. More than 70 platforms already read the format, and nearly 280 million skills have been installed. A skill you write now has a neutral home, kept by engineers from Anthropic, Google and OpenAI. The cocktail of AI-redesigned viruses wiped out the patient's drug-resistant bacteria in lab tests before doctors gave it. The outcome is private, and the company says one case proves little. 29 of 35 answers said people put belonging ahead of truth, and the whole test cost $1.54. Run the prompt on the models you use and see how alike they are. Fable mostly turned letters into the heights of buildings or grass, and a fresh copy decoded 78% from screenshots. The authors say it is not dangerous yet, but the messages were hard to spot by eye. He walks through PHP2Shell, eMudhra's guessable keys and three GitLab 10.0 bugs in about ten days. His take: AI shortens the gap from disclosure to attack, but gets more blame than it earns. He compares Daniel Ellsberg's years as a RAND nuclear planner to today's AI labs, where staff act sure the world changes soon. A short read on how belief shapes behavior inside AI companies. Government websites and public services across Europe went offline this week as organizations responded to two newly disclosed PHP vulnerabilities… · The New Stack How a live GreyNoise product demo turned into an incident response. Inside a proxyjacking campaign that hit 143 devices across 34 countries. · GreyNoise The Python image library's October 16 release patches a JPEG decoder overflow and an EPS flaw that can be chained to run code. If your app opens images from users, upgrade today; it drops Python 3.10. It is fixed in 22.23.4, 24.22.1 and 26.12.1, and OpenAI's Codex Security reported it. Any app that unpacks Brotli data from strangers should update, since Node.js lists no workaround. The k0t crew brute-forced FortiGate logins and used known FortiWeb and GlobalProtect flaws to get in. Take admin pages off the internet, then rotate the LDAP and service passwords stored on each box. In a three-week test, METR found an exploit that could hide hundreds of agent actions and two headers that turned monitoring off. OpenAI says all of it is patched; METR has not checked. Axios reports Morgan Stanley blocks them outright and Wells Fargo is limiting access, and banks would rather give agents read-only access. An agent that pays your bills may soon hit a locked door. A French court ordered the free DNS service to block sports-piracy domains for French users, and Quad9 refuses. If you run offices in France that use it, pick a new resolver before the cutoff. US investors would have to report some deals, and skip others, in AI, chips, quantum and other fields tied to countries of concern. Nothing takes effect until a final rule sets a date. In a note to The Register's columnist, Linus Torvalds said AI only exposes problems that were already there. His complaint is drive-by patches from people who skip learning the code. The BriefAnthropic and a startup called Theorem used Claude to write machine-checked math proofs that 239 test entry points in 123 open-source libraries cannot crash on any input under 64 MB, and the work turned up hundreds of bugs that years of fuzz testing missed. The proofs cover programs built for RISC-V chips under set conditions, so formally verified does not mean bug-free, and only 60 entry points in 25 libraries are public so far. If you maintain an open-source C or C++ library, Theorem will run the same pipeline for free if your project qualifies. Level UpLearn what makes a fuzz test strong enough to prove. Theorem's enrollment guide warns that an assert vanishes in release builds and that returning -1 counts as a pass, so a weak fuzz target turns a proof into an empty promise. Read its checklist and test one of your own fuzz targets against it. →Theorem's OSS Verify enrollment skill for Claude