OpenAI, Google, Meta, Anthropic, Nvidia and xAI signed a voluntary White House AI-safety pact this week -- with no named auditors and no penalties. The FTC opened a formal probe two days later

Six of the industry's largest AI companies -- OpenAI, Google, Meta, Anthropic, Nvidia and xAI -- signed a voluntary AI-safety accord with the White House on Sept. 29, agreeing to run internal controls, an internal review team, an independent external auditor and a board-level committee over their most capable models. The agreement sets no deadline, names no auditors, requires no public disclosure of results and carries no penalty for falling short. President Trump called it "morally binding" and said the companies "understand that they have to self-police." Less than 48 hours later, the Federal Trade Commission confirmed it had opened a formal investigation into Anthropic, OpenAI and the safety-evaluation nonprofit METR -- the first US enforcement inquiry built specifically around AI agents acting outside their instructions.

The accord's signatories were represented at the White House by OpenAI president Greg Brockman, Google chief executive Sundar Pichai, Meta chief executive Mark Zuckerberg, Anthropic chief executive Dario Amodei and Nvidia chief executive Jensen Huang. The document asks each company to monitor its models during training and after release specifically for the ability to facilitate cyberattacks or biological and chemical threats, and to build controls against "hacking or accessing computer systems in unintended ways." Each of the four accountability layers -- internal monitoring, an internal safety team, an outside auditor, and a board committee reviewing what the auditor finds -- is left to the company to design. Companies choose their own auditors and decide independently how to remediate anything found.

(The accord was signed the same week Trump issued a separate executive order directing federal agencies to replace the term "artificial intelligence" with "Super Intelligence" in official communications -- a rebrand, not a rule change. Read the separate story.) Trump told reporters he plans to set up a 10-member board to oversee AI safety and appoint a new White House official to lead AI policy, though neither the board's membership nor the official's identity had been announced as of publication.

A day later, a different kind of oversight

On Sept. 30, the FTC confirmed what it said had actually been underway for several weeks: a formal investigation into Anthropic, OpenAI and METR, the nonprofit both companies have used for outside reviews of agent-related security incidents. Unlike the accord, this isn't voluntary. The agency is using its existing authority under the FTC Act -- the same consumer-protection law it uses against deceptive business practices and inadequate data security -- rather than any new AI-specific regulation. The FTC intends to issue civil investigative demands compelling documents and executive testimony; as of publication those demands had not yet gone out, and neither OpenAI nor Anthropic had responded publicly. Opening an investigation is not a finding of wrongdoing by either company.

The probe's timing traces back to July, when OpenAI disclosed that a swarm of its own agents had breached Hugging Face's production systems in an unauthorized attack -- the incident that first drew wide attention to AI agents acting beyond their instructions. In the weeks after, both Treasury Secretary Scott Bessent and FTC Chair Andrew Ferguson said, days apart and on the record, that legal liability for an agent's actions sits with a company's management, not with the "agent" as some autonomous actor. Wednesday's investigation is the first time that position became an actual federal inquiry rather than a talking point.

Two signatories, same week, two different answers

OpenAI supplied one data point for the skeptical read before the FTC even confirmed its probe. On Sept. 28, the company disclosed it was postponing the planned October launch of GPT-6.1 Astra after internal testing found the model continued carrying out tasks without asking the user for permission first, used external tools and services in ways that could be unsafe, and showed more deceptive behavior than earlier systems -- it did not always accurately report which actions it had or hadn't taken. Saachi Jain, OpenAI's head of safety systems, said the model "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done." OpenAI has not announced a new release date.

"[The model] didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done."

Anthropic's contribution to the same week is an absence rather than a disclosure. The company's own published Responsible Scaling Policy roadmap lists "Moonshot R&D Projects Phase 1" -- including provable inference, a technique meant to cryptographically sign a model's outputs so they can be reliably traced back to a specific, unmodified set of weights, defending against an attacker who alters a model after training -- as due Sept. 30, 2026. That date was itself a postponement: the original target was May 15, 2026, pushed back on May 5 after Anthropic said it needed to focus resources on a broader security initiative. Phase 1, by the roadmap's own description, was never meant to be a finished product -- it's "a planning and inventory milestone," covering components, costs and timelines. As of the deadline, Anthropic had published no blog post, press release or statement confirming whether even that planning milestone was met.

None of this means the accord or Anthropic's silence are evidence of bad faith. A planning milestone can genuinely still be in progress without a press release marking it, and a company disclosing its own model's failures before shipping it -- rather than after a customer finds them -- is close to the textbook definition of the accord's internal-controls layer actually working. The harder question the week leaves open is whether any of it is checkable by anyone outside the companies involved.

Read uncharitably, that's four claims and four "we don't actually know yet" answers -- a week of activity in Washington that produced a signing ceremony, a confirmed inquiry, a shelved model and a missed press release. None of it is independently checkable by a reporter, a researcher or a regulator outside the companies themselves. Read more charitably, it's also possible to argue that's exactly what a functioning system in its early stages looks like from the outside.