On September 30, 2026, the FTC confirmed it is investigating OpenAI, Anthropic and other AI companies over the potential dangers of their products, a day after industry leaders signed a voluntary pledge at the White House. Semafor reports the safety evaluator METR is also a target and that civil investigative demands are expected within weeks.
What should we watch for?
Ackuity's view: a pledge is a statement of intent, not a control. Every agent action still needs an allow, constrain or block decision before it runs, whatever regulators eventually require.
What does the report establish?
This is an investigation, not a finding of wrongdoing. The FTC declined to name the other companies being investigated, and the civil investigative demands had not yet been sent as of the reporting.
Read the full account
On Tuesday, September 29, 2026, executives from AI and tech companies including OpenAI and Anthropic signed a short, voluntary, nonbinding accord at the White House. The next day, an FTC spokesperson confirmed to CNBC that the agency has opened an investigation into OpenAI, Anthropic and other AI companies over the potential dangers of their products. Semafor reported that METR, a nonprofit that evaluates frontier AI models, is also a target, that the probe began before the OpenAI and Hugging Face incident, and that the FTC is set to send civil investigative demands, which are similar to subpoenas, in the next few weeks. ABC News reported that the probe concerns allegations of unfair or deceptive acts and potential harm to consumers.
Voluntary commitments and formal investigations arrived in the same week. For companies deploying AI agents, outside scrutiny of how those agents are controlled is increasing.
Original reporting & evidence
- CNBC: FTC is investigating OpenAI, Anthropic and other AI companies over product risks
- Semafor: FTC probes OpenAI, Anthropic, and METR
- ABC News: FTC opens probe into safety of AI, including Anthropic and OpenAI
How can you check whether an AI safety commitment is implemented?
Turn the stated commitment into specific permitted and prohibited actions for one workflow. Identify the enforcement point, policy owner and evidence required, then run a safe allowed control alongside a denied case. Treat documented intent, configured controls and observed outcomes as different levels of evidence.