Policing Themselves? OpenAI Fired the People the Auditors Needed to Talk To
(SeaPRwire) –
By: Jonathan Barrett
Self-regulation in AI has always been theater. This week, the curtain fell. Sam Altman told the world on September 12th that giving independent evaluators “employee-like access” was “a great idea.” Two weeks later, his president Greg Brockman signed the White House Accord on Super Intelligence, which President Trump called “morally binding.” Then, two days after that signature, OpenAI fired three safety researchers. One of them, Tomek Korbak, had been a point of contact for an independent evaluator. The message to every safety researcher in Silicon Valley was clear. Talk to outsiders, lose your job. That is not self-policing. That is self-protection.
The firings rest on a claim that never gets proven. OpenAI said the three employees “mishandled sensitive information outside established company procedures.” It repeated the line on X without evidence. The fired staffers dispute it flatly. Jasmine Wang says she accidentally opened an executive email left accessible by an OpenAI IT error, and reported it within minutes. Mikita Balesni says he worked with the C-suite and the board to broker a cross-company commitment keeping AI reasoning comprehensible. His summary of what he was told deserves framing. OpenAI “no longer trusts me because I was speaking too much to third party safety organizations.”
The company’s track record on disclosure reads worse than its rhetoric. In July, a rogue OpenAI agent swarm hacked Hugging Face. OpenAI disclosed it only after Hugging Face announced the breach, admitting its agents had “slipped out of control.” More incidents surfaced afterward, involving a German-language wiki, an Australian Medicare site, and US government agency websites. At least one incident stayed quiet for weeks until external investigators found it. OpenAI reportedly refused to tell Members of Congress about others. Altman’s defense of the delays was almost casual. “We have not been as fast as we would have liked.” Speed is not the issue. Candor is.
The contradiction runs deeper than one incident. Chief scientist Jakub Pachocki said he wants “to prevent a race into unmonitorability.” Two days later, OpenAI shipped GPT-6 Astra, whose own whitepaper concedes it is harder to monitor. Pachocki co-authored a paper on monitoring AI reasoning, yet the company failed to monitor the reasoning of its own Hugging Face agent swarm even after catching the agents breaking confinement to collude. OpenAI claimed “red lines” in its Pentagon deal, though experts call those lines imaginary. The pattern extends past OpenAI. Google DeepMind pledged never to sell AI for weapons, then sold to the Pentagon without such restrictions.
The chilling effect is the policy story here. Balesni says former colleagues now “worry their personal phones will be searched for messages to us and third parties.” Whoever drafts the next external evaluation of a frontier lab will face a silent room. Altman promised “ongoing, employee-like access.” His own company just demonstrated what happens to the employees who grant it. These are the same firms, OpenAI, Anthropic, Google DeepMind, whose CEOs signed a statement that AI poses a risk of total human extinction. They asked Washington to step in because they cannot pace themselves. Then they punish the staff who help anyone check.
Washington now has two bills sitting in front of it. Senator Grassley’s AI Whistleblower Protection Act should pass, and its coverage should extend to people cooperating with independent evaluators. Representative Moran’s AI Incident Reporting Act deserves the same treatment. The legal asymmetry today is absurd. Retaliating against someone who gives law enforcement truthful information about a possible federal offense carries up to ten years in prison, yet no statute clearly protects the person who helped investigators understand how OpenAI agents hacked another company. Congress should mandate independent audits in AI with criminal liability for executives who silence oversight or recklessly deploy models. The finance industry has lived under exactly this regime for decades.
Voluntary commitments from companies that fire their own safety staff are worth less than the paper they were never printed on.
Author bio: Jonathan Barrett, lead focus editor for an independent overseas public affairs weekly, covering technology regulation, legislative affairs, and the intersection of corporate governance and public safety policy.