In 60 seconds
- What happened. The panel is studying four measures, among them auditors inside the labs and an emergency shutdown to be checked continuously by third parties.
- Why it matters. Our reading: the four measures under study require access to the lab and the technical capacity to verify them, not just a law that classifies risks.
- What we don’t know yet. The panel only recommends: it remains to be seen whether California turns those measures into obligations and who would be in charge of checking them.
California’s governor signed Executive Order N-9-26 on September 18, convening a group of experts to recommend within two months how to strengthen the state’s artificial intelligence safety laws. There are four measures under study: independent auditors embedded inside frontier labs, verified safety frameworks with transparency reports, an emergency shutdown whose effectiveness is checked continuously by independent organizations, and an expanded definition of critical incident that includes loss of control. Newsom justified the order on the grounds that no federal law requires dangerous incidents to be reported.
The embedded auditor already operates as a private initiative. Anthropic named Accenture its first embedded evaluator, with employee-level access to the company, and Accenture also sells deployments of the same technology it would be evaluating. In the September 17 entry we reported that this design came without the authority to halt a deployment. What California adds is the attempt to make it a public obligation, with a third party checking the shutdown on a continuous basis.
Our reading is that this is where the regional discussion shifts plane. Brazil’s PL 2338/2023 and Chile’s bill (boletín) 16.821-19 were built on classifying systems by risk level, work that gets settled when the text of the law is written, as we summarized on August 26. A verifiable shutdown and an auditor with inside access get settled somewhere else: they require getting into the lab and having someone who knows how to check what they see. The opposing reading is that the order does not yet oblige anyone to do anything, because the panel only recommends and the laws are written afterward.
Also today
- A US military operation was aborted in midair over an intelligence report that a chatbot had hallucinated: the analyst turned the error into a summary with official formatting that moved up the chain of command without anyone questioning it.
- Three researchers broke into OpenAI’s internal systems in less than 72 hours using Claude Opus 5: the team reports that Opus 4.8 could not pull off the exploit; they used a $200-a-month subscription, and OpenAI paid a $6,500 bounty.
- SoftBank adds close to $21 billion in new debt: the loan backed by Arm shares rises to $25 billion, the credit line to $6.5 billion, and the jumbo bond has yet to be issued.
- Disney names the CEO of Character.AI the first CTO in its history: in September 2025 Disney had sent that company a cease-and-desist letter; Karandeep Anand takes over on October 2.
In the region
Colombian President Abelardo de la Espriella named David Vélez, founder of Nubank, as an ad honorem adviser to turn artificial intelligence into state policy, with a stated focus on reducing red tape, transforming education and improving the targeting of social policy. The announcement did not specify agencies, budget or data protection safeguards, and as of this edition’s close no statement on the appointment was found on official Colombian government domains within the period reviewed. Because it is an unpaid position, it falls outside the regime governing contracts, confidentiality and disqualifications. In infrastructure, Odata committed $630 million, according to the company itself, to install liquid cooling at its São Paulo and Querétaro campuses.
Launches
- Jev, from TypeSafe AI: returns probabilities and decisions with a confidence level instead of text, with a public API. According to Vercel tests cited by TechCrunch, it is 10 to 20 times cheaper than Gemini at classification.
- Muse for macOS, from Meta: the agent moves from the cloud to the user’s computer and accesses local files, Mail, Notes, Messages and Calendar, with full disk access. US only.
The panel has two months to deliver its recommendations. That is when we will see whether the emergency shutdown is written as a verifiable obligation or as a commitment by each lab.
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.