US government disables Anthropic models, subpoenas OpenAI in AI crackdown

The US government escalated its oversight of frontier AI this week with two landmark actions. An export-control order at 5:21 PM ET on Friday forced Anthropic to disable its newest models—Fable 5 and Mythos 5—globally, just three days after launch. The trigger: a jailbreak that demonstrated the models could identify software flaws, raising national security concerns. Separately, New York's attorney general served OpenAI with a subpoena covering advertising practices, user engagement, model behavior, consumer/health data handling, and treatment of minors and seniors—signaling consumer-protection law moving into the model layer.

These moves fundamentally reprioritize frontier AI risk. Anthropic had argued for stronger controls; now it faces those controls as an enforcement action. For investors, capability is no longer pure product advantage but regulatory inventory—a model can be state-of-the-art on Monday and policy-frozen by Friday. The security ecosystem saw parallel developments: agentjacking attacks that plant malicious instructions in Sentry error events, LangGraph vulnerabilities enabling RCE in self-hosted agents, and over 400 backdoored AUR packages targeting developer secrets. Mistral eyes a EUR3B raise at a EUR20B valuation as Europe pushes for sovereign AI, while Meta's internal AI unit faces revolt from 1,600 employees over training puzzles.