Top story
Anthropic releases a prototype for proving which model answered
Anthropic released FlashProve on November 2, saying it can verify output against committed model weights without revealing them. The company reports proof generation at 110 tokens per second for quantized Llama 4 Maverick, using roughly 70–95 times the GPU time of vLLM inference. Its research prototype covers integer computation with greedy decoding and contexts of up to 1,024 tokens; formal correctness guarantees apply to its slower Lean verifier, while the fast C++ verifier has been tested for agreement with it.
ThereforeAnthropic says customers and evaluators could check model identity without possessing the weights or rerunning the model themselves.
Anthropic researchCopy link to: Anthropic releases a prototype for proving which model answered
Also
NCSC warns AI can exploit long-tolerated vulnerabilities at scale
ITPro reports that the UK agency's November 3 annual review warns of repeated AI-powered attacks on critical infrastructure over the next two years.
ThereforeThe agency urges organizations to modernize vulnerable technology and use existing security standards.
ITPro · NCSC coverageCopy link to: NCSC warns AI can exploit long-tolerated vulnerabilities at scale
OpenAI plans a Cathedral security sprint and raises bug bounties
Its November 2 announcements describe a planned model-assisted test of its defenses and rewards up to $30,000 for critical core-product flaws, with exceptional findings eligible for up to $172,000 including bonuses.
ThereforePending bounty reports qualify for the new reward ranges, while qualifying awards from the preceding 30 days receive retroactive increases.
OpenAI security sprint · OpenAI bounty changesCopy link to: OpenAI plans a Cathedral security sprint and raises bug bounties
Tools & releases
Apple Watch conversation tools enter developer beta
TechCrunch reports that opt-in Live Rewind and Siri Recaps are rolling out for testing: one displays the previous 15 seconds as text, while the other creates high-level conversation summaries. Apple's documentation warns that important details may be wrong or omitted.
Apple Support · TechCrunchCopy link to: Apple Watch conversation tools enter developer beta
Hugging Face proposes Model Integration Contracts
Its small CPU test models caught a historical UMT5 answer-leakage bug in replay; the checks currently report CI failures for review, alongside existing full-model tests.
Hugging FaceCopy link to: Hugging Face proposes Model Integration Contracts