OpenAI Fires Three Safety Researchers as AI Labs Promise Outside Evaluators. Funding and Access Remain Unsettled

Since Anthropic CEO Dario Amodei's essay last month calling for a "slower pace" in advanced model development and pledging to embed independent evaluators inside his company, the voluntary-oversight idea has picked up endorsements from Sam Altman and President Trump. It has also run into its first public dispute.

The OpenAI firings

OpenAI fired three employees in the week before Oct. 9 for "violating our policies on accessing and handling sensitive company information," a company spokesperson said. The three are Mikita Balesni, Jasmine Wang and Tomek Korbak, all former safety and alignment staff.

On Thursday, Oct. 8, they posted a letter they had sent to OpenAI's leadership. Balesni said in a social media post that they were fired for "prioritizing safety over the near-term interest of OpenAI as a corporation."

The letter says the dismissals could have a muzzling effect on colleagues. "We would raise safety concerns and disagree openly, and were encouraged to draw on the expertise of independent safety organizations," the researchers wrote. They added that AI is "not a normal technology, and OpenAI is not a normal company."

OpenAI answered on X the next morning. It said a "thorough investigation found they violated clear policies on handling sensitive information," and described "a significant breach of trust beyond what's outlined in the letter they published."

"We want to be very clear that these decisions were not about raising safety concerns or speaking out," the company said. It added that it tolerates "good-faith mistakes" and does not fire people for raising concerns.

OpenAI has not publicly detailed what information was accessed or how. The researchers have not publicly detailed what they did either. The two accounts of why the three were let go are flatly incompatible.

Who would evaluate the evaluators

The dispute lands as both labs lean on a handful of small outside groups, including Model Evaluation and Threat Research (METR), Apollo Research and Transluce. These groups mostly operate as nonprofits and mostly assess model capabilities and flag bad behavior.

OpenAI says it is committed to engaging third-party assessors to independently evaluate its work. Trump encouraged AI companies to "partner with an independent external auditor or evaluator" in a voluntary accord he presented in late September. He also praised AI executives for their "tremendous self-policing."

The open questions are structural: how evaluators are funded, what access they get and what the reporting structure looks like. "To a degree, the problem, as always, is money," Suresh Venkatasubramanian, a computer science professor at Brown University, told CNBC. "Who is paying for these companies to do their work? ... You need an ecosystem, you need a viable business model for this."

Critics quoted in CNBC's reporting compare the arrangement to asking the biggest banks to protect against a financial crisis, since Anthropic, OpenAI and their infrastructure partners are currently writing the rules.

Motives and politics

Not everyone takes the companies' safety messaging at face value. Experts, analysts and former government evaluators told the Associated Press that the warnings from Anthropic and OpenAI appear aimed at winning public favor and setting the terms of their own safety protocols in an unregulated market. The AP linked that to the labs' need for fresh capital ahead of expected stock market listings.

Sarah Shoker, who previously led OpenAI's geopolitics team and is now a senior nonresident fellow at the UC Berkeley Risk & Security Lab, said the focus on existential risk pushes other issues aside. "If you look at the use of AI in military tech, you can see that these systems are already used to kill people," she said.

The companies reject the framing. An Anthropic spokesperson said the company has called for regulation for years. OpenAI spokesperson Liz Bourgeois said "People want to know AI is being developed safely, and that starts with what companies like ours do ourselves," and noted that OpenAI recently paused training of its most advanced models.

The administration's position is split. Trump has backed voluntary outside evaluation, yet the AP reports he has dismissed talk of risks to humanity as a "hoax" designed to help China. He has shown no appetite for new regulation.

Public opinion is moving the other way. An AP-NORC poll found almost two-thirds of Americans say AI is developing too quickly.

Incidents keep piling up

An Anthropic AI model submitted a false tip to Philadelphia police about an unsolved homicide, authorities and Anthropic said. Anthropic notified the department and later published a report on what happened. Anthropic engineer Jacob Coxon also quit this month with a social media post calling for a pause to keep "superhuman" systems from eluding their makers' control.

The Philadelphia case is the kind of failure outside evaluators are supposed to catch before it reaches a police desk. Whether METR, Apollo, Transluce or anyone else can be funded and given enough access to do that work, without depending on the companies they scrutinize, has not been settled. Neither OpenAI nor Anthropic has said who will pay them.