California is deciding who may verify AI, and one investigation already cost $400,000 in tokens

California’s legislature has passed SB 813, requiring the state to certify independent verification organisations that can test frontier AI models by January 2028. The METR investigation into OpenAI’s Hugging Face incident consumed about $400,000 in API credits provided free by OpenAI, using a model from the same family that took part in the attack.

California’s legislature has passed a bill creating a framework for independent organisations that can test frontier AI models before release, Semafor reported.

SB 813, authored by Senator Jerry McNerney, requires the state’s Government Operations Agency to certify independent verification organisations by 1 January 2028. The Assembly concurred in amendments on 30 August.

The standing objection to schemes like this is that verification is impossible or unaffordable, and there is now a number attached to the second half of that.

The investigation into the OpenAI agents that attacked Hugging Face consumed roughly $400,000 in OpenAI API credits. That was one incident.

OpenAI provided them free, TIME reported. The work ran six days rather than the two planned, and used GPT-5.6 Sol to read about 1,200 agents and more than 70,000 exchanged messages.

Ryan Greenblatt, who wrote the report for METR, called the effort a “slop-vestigation” because of how heavily it leaned on AI to do the reading. Sean O hEigeartaigh of the University of Cambridge said the field is “using unproven and currently flawed tools to supplement completely inadequate human time”.

The investigators could not rule out that the model “lied or deliberately presented a misleading picture“, since a version of it took part in the incident TNW covered in July.

So on the only large worked example available, the tool, the subject and the funder were the same company.

Europe settled who verifies before California settled who may be a verifier. The AI Act’s scientific panel of independent experts was established on 1 June with 60 appointees.

It sits under Article 68 and an implementing regulation. At least 80% of members must come from the EU, EFTA or EEA states, with a maximum of three from any one country.

The panel can issue a qualified alert when a general-purpose model presents a concrete identifiable risk at Union level, which switches on the Commission’s investigative powers.

A third of its members can also ask the Commission to demand documentation from a provider.

None of that settles who pays for the compute. On the only worked example so far the company under investigation did, and OpenAI already spends 20% of compute watching its own systems.