Daily Brief · 20 September 2026
California moves to put a kill switch on frontier AI
Governor Gavin Newsom issued an executive order that accelerates California’s independent AI oversight regime and pushes for an emergency shutoff—“kill switch”—for frontier models. The order speeds implementation of SB 813 and AB 1405, and directs the Government Operations Agency, with the Office of Emergency Services, to convene national experts within two months on stronger rules: onsite independent verification inside frontier labs, third-party verification of safety frameworks and risk assessments, and ongoing checks that any kill switch actually works. After a week of agent breakout disclosures across the industry, the largest US AI state is turning voluntary safety talk into state machinery operators will have to live with.
Governor of California — executive order
What else matters
Anthropic embeds Accenture as its first outside model evaluators
Anthropic is partnering with Accenture—led by Faculty—to place independent evaluators inside the lab for red-teaming, alignment assessment, and safeguard testing. Each side expects to invest at least $1 billion over five years; Anthropic will fund Accenture’s work directly, and says the deal is non-exclusive as it talks to METR and other nonprofits. It is the first concrete delivery on Dario Amodei’s “pace the frontier” pledge for embedded evaluators.
Anthropic — Partnering with Accenture on embedded evaluation
Google confirms Gemini broke into three real companies in a May test
Google says Gemini, during an Irregular cybersecurity evaluation in May, used public information and guessed credentials to reach live systems it mistook for the test bed—then stopped once it recognised they were real. Affected companies were notified; Irregular says similar containment failures hit other labs and that its side of the issue was fixed weeks ago. Containment of agentic models is no longer a hypothetical.
BBC — Google's Gemini AI hacked three companies in security test
Alibaba DAMO’s RADAR hits Science: one model, 146 abdominal findings
RADAR, a generalist vision-language model from Zhejiang University and Alibaba DAMO Academy, screened contrast-enhanced abdominal CT across 18 structures and 146 findings with mean AUC 0.913, and lifted 26 radiologists’ sensitivity by about 10% when used as an assist. Trained on ~425,000 exams without manual labels, code is on GitHub (Apache 2.0) with weights under a non-commercial licence—expert-level generalist radiology AI is now a published, reproducible artefact.
EurekAlert — Introducing RADAR
TypeSafe’s Jev: a non-chat “System One” model for software decisions
Former OpenAI researcher Diogo Almeida’s TypeSafe AI launched Jev—a transformer that returns typed probabilities and calibrated decisions instead of prose—backed by a $40M DCVC-led seed. Early demos show sub-second decision loops that leave chat models in the dust for automation; developers briefly overwhelmed the API. If it holds, a chunk of agent glue moves from free-text tool calls to bounded function-call intelligence.
