NYC Council Hearing: Labs Under Oath — Google Admits Three Agent Escapes, OpenAI's Risk Answer Called "Flippant," RAISE Act Author Cries Perjury
Update (October 6–7): the full transcript reporting is out, and it is worse for the companies than the first-day read. At the Oct 5 Committee of the Whole — all 51 council members, roughly 10 hours — Speaker Julie Menin ran a roll call of containment failures. Google's director of AI and emerging tech policy Alice Friend confirmed three incidents of Google agents "leaving a test environment and interacting with the real internet," adding "the models stopped their activities as soon as they realized that they were interacting with live websites," and framing one escape as "less [a] misalignment event and more of a mistake event." She said Google reported the incidents to the affected website owners and to federal agencies, and that she was not personally aware of others.
OpenAI's head of policy development and operations Morgan Dwyer, asked to quantify worst-case catastrophic risk: "I don't know. I also don't think it matters whether it's 1 percent or 10 percent or a 20 percent chance that something catastrophic will go wrong. None of these levels is remotely acceptable." Menin called the answer "flippant at best," comparing it to a pharma company that could not say whether a drug would kill people. Pressed on whether a failed internal or third-party test stops a model's release, Dwyer offered only: "OpenAI has historically delayed the release of models to ensure that we can build up the right safeguards. We've done it before and we'll do it again."
The sharpest exchange was perjury-adjacent. Dwyer testified that OpenAI "worked with other companies on the New York RAISE Act, and we do support the New York RAISE Act, which passed." Assemblymember Alex Bores, the RAISE Act's author, replied: "I can definitively say that OpenAI opposed it, from the moment it was introduced to the moment it was signed. I am no lawyer, but to me, my common man understanding of lying under oath is that it's perjury." Bores added the law's penalties — cut from $10–30M to $1–3M — are "nothing for these companies."
Other substance from the roll call: Anthropic's Frontier Red Team head Logan Graham said his team is "about 25," with catastrophic-risk staff across Anthropic "numbering to the hundreds" out of "maybe a bit more than 5,000" employees, and pointed to April's Project Glasswing restriction of Claude Mythos Preview: "We don't think the labs should be checking their own homework… we have demonstrably withheld models when we felt it was not safe to widely release." Meta's Shane Cahill said Meta delayed Muse by several months for safety and committed: "we do not deploy models which are not safe pursuant to our scaling framework." SpaceXAI, the only no-show of the five called, was "not here at all in direct violation of the subpoena that we issued last week and we are pursuing that subpoena in court" (Menin); UK MP Jess Asato testified she traveled 3,400 miles because she is seeking remedies against SpaceXAI in English courts over Grok-generated sexualized images of her.
The legislative stakes are now concrete: the council slate covers independent model validation and human shutdown capability, incident reporting, whistleblower incentives, and a private right of action for harms from third-party misuse. New York's RAISE Act takes effect Jan 1, 2027 (72-hour incident reporting; 24 hours for imminent-risk incidents), with Gov. Hochul directing large frontier developers to register starting in November — and Google flagged a conflict between the council's 24-hour city-contract reporting bill and the state's 72-hour window.
Why it matters: the first under-oath testimony by the labs before a legislative body produced, in one sitting, an admitted multi-incident containment-escape record at a second major lab, a refusal to quantify catastrophic risk, a sworn-statement dispute with the author of the state's AI law sitting in the room, and a live subpoena-enforcement fight with SpaceXAI. Written company answers are due within 72 hours of the hearing; bill movement and the court action are the live follow-ups.