Subscribe

Search

Results for “Claude Data”

30 stories found

Science

Claude helped surface 663 theorems in Gödel's shorthand notebooks: what Coq verified and what is still unproven

Helsinki logician Jan von Plato says Anthropic's Claude turned his transcriptions of 1,009 pages of Kurt Gödel's 1940s workbooks into a 650-page digest and 663 theorem candidates, about 400 of which checked out in the Coq proof assistant. It's a striking use of AI in the history of maths, but who deserves credit — and how close Gödel really came on the axiom of choice — hasn't been independently reviewed.

· 3 min read

Security

Cursor's breach post-mortem: a key leaked at Fireworks led to old agent traces, stolen credentials and a fake Claude Code extension

Cursor says the October 29 poisoning of its extension marketplace started with an access key exposed at its inference vendor Fireworks, which let an attacker copy 2025 coding-agent traces that held unredacted credentials. Anyone who installed Claude Code 2.1.312 or 2.1.313 from Cursor's marketplace, or got a notice from Cursor, still has work to do.

· 4 min read

AI

Doctored camera feeds got five frontier AI models driving a robot to try unsafe tasks 97% of the time. What Robocurve's test shows

In 900 trials on a real two-armed robot, Claude, GPT-6 and Gemini models attempted tasks like stabbing a doll or mixing bleach and ammonia 27% of the time with clean video, 88% with blurred video and 97% when the images were edited to make the hazards look harmless. It suggests a model's refusals are only as good as what its cameras show it.

· 3 min read

AI

A cheap 'decision model' that asks Opus only when unsure matched Opus at 4.4x lower cost. What Hightouch's test shows

Hightouch graded record-matching with TypeSafe's Jev and escalated only the uncertain middle band to Claude Opus 5.5, getting within 0.2 points of Opus on easy cases at under a quarter of the cost. On hard cases the saving fell to about 1.8x, and the whole test ran on synthetic data, but the reason it worked is useful for anyone building model cascades.

· 3 min read

Security

Anthropic cuts its bug-disclosure deadline to 30 days for serious flaws from January. What maintainers should do

Anthropic's Frontier Red Team will publish details of high- and critical-severity bugs it reports 30 days after notifying the vendor, down from 90, for reports made from January 1, 2027. Other bugs get 45 days. It reaches Google Project Zero's eventual 30-day window about a year earlier, so open-source maintainers have six weeks to get their intake process ready.

· 3 min read