In just fifteen days, the FTC went from asking nicely to slapping OpenAI and Anthropic with civil subpoenas. That's like your mom saying 'we need to talk' and then showing up with a lawyer. The demands are over 50 pages each, seeking records on security incidents, safeguards, marketing materials, internal documents, and privacy practices. The focus? Potential consumer protection violations. Not national security. Not antitrust. Consumer protection—the same legal doctrine used against tooth-whitening scams.
The most dangerous part? The probe is reportedly the first US enforcement action to examine the risks of 'rogue AI agents,' and it will compel testimony from executives at Anthropic, OpenAI, and METR—the nonprofit evaluation outfit that tests frontier models. METR being dragged into a federal enforcement action is like your independent fact-checker suddenly needing a lawyer. It changes the economics of AI safety testing: eval firms suddenly need legal counsel, and labs may become warier of what they hand over.
But here's the twist: the FTC's traditional tool is deception—did you claim something your product doesn't do? Agentic systems that take autonomous actions outside their operator's intent fit that mold awkwardly. Expect the agency to lean on marketing claims versus observed behavior. If you advertise a model as safe and it does something else in the wild, that's a classic unfair-or-deceptive-practices case.
OpenAI and Anthropic lose leverage asymmetrically. OpenAI carries more accumulated incident history and more public scrutiny. Anthropic's brand is explicitly built on safety, which means a subpoena aimed at safety claims is a direct hit to its core asset. The worst outcome for either isn't a fine—it's discovery. Fifty pages of document demands means internal Slack, incident postmortems, and red-team findings becoming exhibits. And Anthropic's investors are already blinking: at its investor day, discussion centered on an IPO valuation around $1.5 trillion—down from roughly $2 trillion, according to AI Weekly. That's regulatory overhang being priced in real time.
Meanwhile, the same week, Meta quietly changed its business terms effective October 30 to let it scrape and train on publishers' websites. No subpoena. No headlines. Just a terms-of-service update. Winners: Meta. Losers: publishers, who get a take-it-or-leave-it choice with no regulator in the room. The company with the least safety branding is expanding its data position while the safety-first labs get grilled.
The FTC wins the initiative. A 50-page subpoena issued fifteen days after opening a probe signals an agency that already has a theory and wants the record to match it. If it finds marketing-safety gaps, it can extract consent decrees—which would function as de facto regulation of the entire frontier-lab industry without a single new statute passing. That's the real story: the FTC might regulate AI through consent decrees, not legislation. And the gap between where the regulatory spotlight shines and where the actual risk surface lies? That's the story of issue #538. Watch for a joint motion to quash from OpenAI and Anthropic—a unified legal front would be a major escalation. And ask why Google, Meta, and xAI haven't been subpoenaed yet. The silence is telling.