The news
Independent AI audits moved from proposal to practice in September 2026. Between September 9 and 23, Anthropic and OpenAI agreed to let outside evaluators work inside their labs, and California enacted a system to certify and register AI auditors.
The trigger was a post by Anthropic chief executive Dario Amodei titled "We must pace the frontier." Reuters' timeline dates the post about September 12. He argued the industry should slow how quickly it improves model capabilities, without halting, and committed Anthropic to hosting outside review teams with desks, badges and access mostly comparable to its own risk staff. Those evaluators could publish key findings without Anthropic's editorial control, though Anthropic could redact security-sensitive, legally privileged, commercially sensitive or third-party confidential information. He also asked the US government to facilitate industry safety talks.
OpenAI chief executive Sam Altman said he agreed on pacing and that OpenAI would adopt evaluators with similar access, MediaNama reported. Elon Musk also endorsed the call, according to Reuters. Meta Platforms (META) chief executive Mark Zuckerberg rejected a coordinated slowdown on September 15, writing on X that every lab has its own responsibility and incentive to train models safely, Reuters reported.
Anthropic moved first from commitment to contract. On September 18, Anthropic said Faculty, the specialist AI business of Accenture (ACN), will evaluate and red-team its models with access comparable to an employee's, funded by Anthropic. Anthropic said each company expects to invest at least $1 billion in building evaluation capacity over the next five years. OpenAI published principles for third-party assessments, Resultsense reported on September 23, covering safety cases, safeguard testing, dangerous-capability evaluations and investigations of serious misalignment incidents. Resultsense reported that assessors must disclose conflicts, including payment arrangements, and keep editorial independence, though OpenAI agrees the scope, sees findings first and can request redactions, which Resultsense likened to a commissioned audit.
California moved in parallel. On September 9, Governor Gavin Newsom signed SB 813, which creates a framework for certifying independent verification organizations, and AB 1405, a state registry for AI auditors with independence standards, StateScoop reported. On September 18, Newsom issued an executive order to speed both laws and have experts recommend, within two months, possible changes to state law, such as requiring onsite verification organizations at frontier labs and an emergency shutoff, or kill switch, for frontier models.
The numbers
- Anthropic and Accenture expected investment in evaluation capacity over five years
- At least $1 billion each
- California AI oversight laws signed September 9
- 2 (SB 813 and AB 1405)
- Deadline for California expert group recommendations
- Within two months of September 18
- Priority areas in OpenAI's third-party assessment principles
- 4
Why CEOs should care
For boards and risk committees, AI vendor due diligence can now ask for evidence rather than assurances. Ask each model provider whether it hosts independent evaluators, who pays them, what access they have and whether their findings are published. OpenAI's own principles call for disclosure of payment arrangements, which is a reasonable standard to apply to any vendor's safety claims. Add these answers to vendor risk reviews and renewal decisions.
General counsels and compliance leaders should expect scrutiny to spread from model makers to the companies that deploy them. Newsom called California's approach a model that should become the national baseline, and StateScoop noted Illinois passed a law in July requiring annual independent audits of major frontier developers. Keep documentation on how your own AI systems are tested, monitored and approved so that an outside reviewer could follow it.
CISOs and technology leaders should read California's kill-switch work as a signal of where requirements could head; the state's experts will also consider adding loss-of-control incidents to the definition of critical safety incidents. Any agent your company runs should have a tested way to stop it, logs that show what it did, and a named owner who can make that call.
The bigger picture
The industry is splitting on how to govern frontier AI. Anthropic and OpenAI favor pacing plus embedded outside checks and Musk has endorsed pacing, while Zuckerberg argues each lab should set its own safe pace. Pacing has limits in practice: Anthropic and OpenAI both released new models on September 22. For buyers, the shift that matters is that safety claims are becoming auditable by third parties, which gives customers a new basis for comparing providers beyond benchmarks and price.
What’s next
Watch for California's expert group recommendations on independent oversight and the kill switch, due about two months after the September 18 order, and for the additional evaluators Anthropic said it would name in the coming weeks. Whether Meta, Google or other developers accept embedded evaluators will show whether this becomes an industry norm or a split market.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made with our newsroom’s technology tools, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is not published automatically; it is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
