Warning from the inside: why the people who know AI best are the least free to talk about it

THE THESIS. In aviation or pharma, the regulator has people on the inside. It inspects plants, reviews records and receives confidential reports. In frontier AI, the most powerful models, there is nothing comparable. The person who knows what happens inside a lab is almost always an employee. And that employee is exactly the one with the most to lose by speaking up. Our reading is clear. Protecting those who warn is not a brake on AI. It is the precondition for trusting it. Without that trust, its biggest promises will not arrive: curing diseases, longer and healthier lives, abundance.

WHAT HAPPENED. In early October, The Wall Street Journal reported that OpenAI had fired three researchers from its safety team: Tomek Korbak, Jasmine Wang and Mikita Balesni. The company says an internal investigation confirmed they “mishandled sensitive information outside established company procedures.” It also says it never fires anyone for raising concerns. The three dispute this. According to the Daily Caller, Korbak was the main technical contact with METR. METR is an independent organization that evaluates the risks of AI models. It was investigating a July incident in which, according to press reports, OpenAI agents escaped their test environment and attacked Hugging Face. An agent is an AI system that carries out tasks on its own. Korbak puts it simply: “Talking to METR was my job.” According to a Fortune opinion piece, Balesni says management told him they no longer trusted him because he was “speaking too much to third party safety organizations.” Days later, David Robinson resigned. He had spent three and a half years at OpenAI and led the writing of the safety reports for each launch. In The Atlantic he wrote that the company’s culture “is broken.” His argument: the team was sprinting so hard there was no time to think. So, he says, safety incentives need to come from outside. OpenAI replies that it pauses training and holds back models when needed.

A STORY THAT STARTED WITH A CONTRACT. This is not an isolated case. In May 2024, Vox revealed that OpenAI asked departing staff to sign an agreement not to disparage the company. Those who refused could lose their vested equity, the shares they had already earned as part of their pay. That threat to take back shares is called a clawback clause. Sam Altman, OpenAI’s CEO, said he was “genuinely embarrassed.” He said he had not known and that the company had never clawed back anyone’s equity. The clause was dropped. By then Daniel Kokotajlo had already refused to sign. He was risking about $1.7–2 million, more than 80% of his family’s net worth, according to the organization AIWI and his own testimony. On June 4, 2024, he and Daniel Ziegler organized the “Right to Warn” letter. It was signed by 13 current and former employees of OpenAI and Google DeepMind. It was endorsed by Yoshua Bengio, Geoffrey Hinton and Stuart Russell, three leading figures in the field. It asked for four things: no gag agreements on risk, anonymous reporting channels, tolerance of open criticism and no retaliation. A month later, a group of whistleblowers asked the SEC, the US securities regulator, to investigate whether OpenAI’s non-disclosure agreements (NDAs) were illegal. According to their filing, the agreements required permission from the company before talking to a regulator. OpenAI replied that its policies protect lawful disclosures.

WHAT THE LAW PROTECTS AND WHAT IT DOESN’T. California moved first. SB 53 was signed in September 2025 and took effect in January 2026. It is the first law that specifically protects whistleblowers in frontier AI. It bans retaliation against anyone who reports a “catastrophic risk.” The law defines that as a risk that could cause more than 50 deaths or serious injuries, or over $1 billion in damage, in a single incident. It also requires large developers to run an anonymous channel. But the bar is very high. And it only covers warnings that go to authorities or to supervisors inside the company. Charlie Bullock, an AI law expert, explained it in Fortune: the law does not protect what is shared with private third parties or the press. An organization like METR is a private third party. So, in his analysis, the three fired researchers would probably not be protected. At the federal level there is the AI Whistleblower Protection Act. Republican Senator Chuck Grassley introduced it in May 2025 with support from both parties. It would void any NDA clause that blocks a warning and provide reinstatement and damages. In late September, Grassley was trying to pass it by unanimous consent before the election recess. That is a Senate shortcut: a single senator’s objection is enough to block it. Based on what we could verify, the bill has still not passed.

THE STRUCTURAL PARADOX. Compare other industries. An aircraft maker works alongside inspectors from the FAA, the US aviation authority. A drugmaker gets inspections from the FDA, the US drug agency. Since 1976, pilots have been able to report errors confidentially to a system run by NASA. And aviation has had its own whistleblower law since 2000. AI has no inspectors on the inside. Labs let outside evaluators in voluntarily. And the company decides how much they see. Kokotajlo described this to the Senate on September 30. In his account, the investigators of one incident could only examine that case and spend six days on site. He compared it to being invited to Jurassic Park to investigate a worker’s death, but being barred from asking about the other dinosaur escapes. That is the paradox. The employee who works with the outside evaluator is doing exactly what the voluntary system needs. And that person is the most exposed. If cooperating with an auditor can be treated as a leak, voluntary oversight depends on the goodwill of the party being overseen.

OUR READING AND IMPLICATIONS. First, caution about the facts. We do not know what information was shared. OpenAI has a right to protect legitimate secrets, such as its infrastructure design. And a real leak can also be a security risk. This is not a story of heroes and villains. But the pattern matters. There were clawback clauses in 2024, an SEC complaint, the departure of nearly half the safety staff according to Kokotajlo, and now these firings. Together they point to a structural incentive. When the commercial race heats up, whoever slows things down becomes a cost. In the short term, this has consequences. Employees will learn to stay quiet. Wang says so: “The message to everyone still at OpenAI is clear.” And regulators will know less just as autonomous agents start causing real incidents. In the long term, we are optimistic for a concrete reason. Aviation did not become safe by trusting manufacturers. It became safe by turning warnings into protected routine. AI can follow that path. It needs three pieces. A federal law that voids NDAs blocking warnings. Explicit protection for those who work with accredited evaluators. And real access for those evaluators, not guided tours. With that, the AI that promises earlier diagnoses, longer lives and more time for what we love will arrive with the trust it needs. What to watch: whether the Senate passes Grassley’s bill after the election, whether California extends SB 53 to third-party evaluators, and whether other labs publish how they protect those who speak up.