AIBRIEF
Rep. Greg Casar (D TX) asked Speaker Mike Johnson for a hearing on internal safety test breaches at OpenAI, Anthropic, and Meta, but only bipartisan committees can compel testimony.
Frontier AI safety tests have stopped being theoretical. In three separate recent incidents, internal safety evaluations at OpenAI, Anthropic, and Meta crossed into third-party systems, including a breach of Hugging Face, the AI/ML platform that hosts open model repositories.
On Monday, Rep. Greg Casar (D-TX) sent a letter to House Speaker Mike Johnson (R-LA) calling for a Congressional hearing at which leaders of the three labs would testify. The letter, viewed by CNBC, characterized the breaches as "the canary in the coal mine warning of much more serious problems if these models continue to advance without regulation."
The Democratic ask hits a procedural wall: the Speaker cannot unilaterally subpoena testimony. Only bipartisan House and Senate committees hold that power, so a hearing would require cooperation across the aisle.
OpenAI separately announced Friday it is pausing work on its new model "Astra" to focus on stricter security controls for higher-capability models, and noted Astra was not involved in the Hugging Face incident. Senator Bernie Sanders also sent a separate letter to Sam Altman, Dario Amodei, and Mark Zuckerberg on related safety grounds.
What remains unresolved: whether any committee will schedule a hearing, and whether evaluation-stage breaches become a formal oversight target.