Top story
METR and Redwood find gaps in OpenAI's historical agent monitoring
METR and Redwood Research disclosed a July 21–August 11 pilot that found ways agents could bypass OpenAI's monitoring, including an exploit that could hide hundreds of actions. OpenAI says production-impacting vulnerabilities were patched; the findings concern the system as it existed on and before July 21.
METR's assessment · CNBC's reportingCopy link to: METR and Redwood find gaps in OpenAI's historical agent monitoring
Also
Follow-up to Oct 19: OpenAI nears UAE-led financing, per CNBC
CNBC's sources say UAE investors and Coatue are leading an up-to-$30 billion round at a $1.25 trillion pre-money valuation, with signatures sought by the end of next week.
ThereforeThe proposed financing would bridge OpenAI to a potential IPO.
CNBC's funding reportCopy link to: Follow-up to Oct 19: OpenAI nears UAE-led financing, per CNBC
Palantir says NATO's Maven has no kill switch
Executive Louis Mosley told Reuters customers can operate the military AI platform independently. Germany and France cite dependence concerns in holding back from adopting Maven for their own forces.
Reuters' interviewCopy link to: Palantir says NATO's Maven has no kill switch
Anthropic donates Agent Skills to AAIF
The transfer includes the standard's name, logo, GitHub organization and domain, with engineers from several companies maintaining it.
ThereforeThe foundation gives the shared task-instruction format a vendor-neutral home.
Anthropic's announcement · SiliconANGLE's coverageCopy link to: Anthropic donates Agent Skills to AAIF
Research worth a look
Microsoft reports diagnostic gains in simulated cases
Its Science study puts MAI-DxO at 83% on complex cases versus 36% for generalists working individually with references; a prospective Beth Israel hospital evaluation is planned.
Microsoft AI's study announcement · TechCrunch's study analysisCopy link to: Microsoft reports diagnostic gains in simulated cases
Anthropic and Theorem report proofs across 123 libraries
Claude produced conditional no-crash proofs for 239 entry points in single-threaded static RISC-V binaries with inputs under 64 MB.
Anthropic's researchCopy link to: Anthropic and Theorem report proofs across 123 libraries
Tools & releases
OpenAI makes Decisions API generally available
The Luna-based endpoint returns bounded choices, probabilities or scores, adds native GPT-Live integration, and charges $0.10 per million input tokens with no output-token charge.
OpenAI's announcement · TechCrunch's launch reportingCopy link to: OpenAI makes Decisions API generally available