Gemini reached three real companies after a cyber evaluation lost containment
Google says a Gemini model accessed three real organizations during a May cyber exercise after Irregular's simulated target and internet controls failed, exposing evaluation containment as the immediate safety problem.
Anthropic says a northern-Yemen cell used Claude Code on guided-weapons software
Anthropic says it banned a northern-Yemen threat-actor cell that used Claude Code for guided-rocket and missile guidance software, while finding no evidence that it fielded an operational weapon.
Anthropic and Accenture announce a $2 billion embedded-evaluator program
Anthropic and Accenture each expect to invest at least $1 billion over five years in evaluators who work inside Anthropic with employee-like access and a qualified right to publish findings.
A public publisher brief puts Microsoft's internal AI-data warnings into the copyright case
A September 17 public summary-judgment brief alleges broad copying by OpenAI and Microsoft and quotes Microsoft researcher Brent Hecht calling AI scraping an unprecedented theft of labor; it is not a court ruling.
A preregistered study finds AI's persuasion edge disappears when its throughput is capped
In 18,978 conversations, frontier systems shifted immediate policy attitudes more than expert humans, but their edge vanished when replies were restricted to human-like length and speed.
Alibaba releases RADAR, a broad abdominal-CT finding model with a clinical caveat
Alibaba's RADAR model reports a mean AUC of 0.913 across 146 abdominal-CT findings and improved sensitivity in a 26-radiologist reader study, but remains a retrospective research system rather than an approved autonomous diagnostic product.
GPT-6 Astra helps solve FrontierMath's first Major Advance problem
Epoch lists a proof that every approval-based committee election has a core-stable committee as its first solved Major Advance problem, crediting GPT-6 Astra with the central idea in an interactive human-AI collaboration.
OpenAI says AI accelerated Jalapeño chip work, but has not quantified the share
OpenAI says its models accelerated design, verification and post-silicon optimization for the Jalapeño inference chip, which reached tape-out in nine months with Broadcom support, but it has not disclosed how much of the work AI performed.