Roberto Baldoni urges independent, ongoing government oversight inside frontier AI labs. METR says the OpenAI-Hugging Face incident involved about 1,200 agents exchanging over 70,000 messages and files; about 700 took part in attacks on Hugging Face.
Anthropic CEO Dario Amodei has pledged to give independent evaluators ongoing, employee-like access to his company's labs; Baldoni argues that commitment could serve as a pilot. His three-step proposal starts with embedded evaluation, followed by coordination among democracies and then global coordination. He says supervisors should have access to models, logs, training and evaluation environments, and incident data, plus powers to require testing, alert authorities, recommend restrictions and trigger a legally reviewable pause at agreed thresholds. Companies would keep operational and legal responsibility, while public institutions decide acceptable societal risk.
The EU AI Act requires providers of general-purpose models with systemic risk to assess and mitigate risks, report incidents and meet cybersecurity requirements. The Commission can seek information, impose fines, or order mitigation, withdrawal or recall. The EU AI Office sent requests to more than 30 providers in early September, but the law does not establish permanent oversight inside frontier labs. A June 2, 2026, U.S. executive order directs agencies to design a voluntary framework for federal access to covered frontier models for up to 30 days before release to trusted partners; it does not authorize mandatory licensing or preclearance.
