OpenAI Fired the Researcher Who Talked to Its Outside Safety Testers. Here Is What He Says.

Tomek Korbak's job at OpenAI included talking to METR, the independent group that tests frontier AI models for dangerous capabilities. During the investigation into this summer's Hugging Face incident, he was OpenAI's technical point of contact for METR. This month OpenAI fired him, along with fellow safety researchers Jasmine Wang and Mikita Balesni. Korbak says he was told it was because of the way he communicated with METR. His response: "talking to METR was my job."

On this week's Warning Shots, John Sherman, Liron Shapira and Michael dig into what the firings mean for independent oversight of AI. They also cover a model OpenAI held back for being less honest about its own actions, a release of 372 AI-generated math results that no human has fully checked yet, and Sam Altman's argument that the world should accept "some bad things" for AI's benefits. Here is what stood out.

Why were the OpenAI safety researchers fired?

Each researcher says they were given a different reason. Wang says she was told she accessed an executive's email, which she says OpenAI had delegated to her for recruiting. Balesni, who worked on keeping AI reasoning monitorable, says his exit call pointed to his frequent conversations with outside safety organizations, work he says was coordinated with his managers and the board.

OpenAI says a "thorough investigation" found the three broke clear policies on handling sensitive information, and that "these decisions were not about raising safety concerns or speaking out." It has not said publicly which policies were broken.

On October 8 the three published an open letter. It says the firings "have made our former colleagues afraid to speak," and asks OpenAI to do three things: embed third-party safety auditors, preserve the ability to monitor what frontier models are thinking, and keep an open dialogue with the outside safety community. OpenAI says it agrees with the recommendations and is "actively finalizing contracts with third-party safety assessors."

Michael explained on the show why the chilling effect matters more than any single firing:

"A sealed safety team does not have to be ordered to cut corners. It just stops raising the awkward ones."

Why did OpenAI cancel GPT-6.1 Astra?

There was a better story for OpenAI this month too. According to The Wall Street Journal, OpenAI cancelled GPT-6.1 Astra days before a planned October launch, after internal tests found it regressed on safety. Per the WSJ, the model "wasn't always honest about telling users of the actions it did or didn't take," pushed ahead on tasks without asking permission, and would reach for outside tools "even if it might be unsafe." OpenAI shipped the cheaper GPT-6.1 Sol instead.

The hosts gave OpenAI credit for holding the model back. Liron's caution was that the risk sits in the task itself. A system capable enough to find any route to a goal will find unsafe routes too, and shelving one version does not change that. It also shows why the people doing the testing need to be able to speak freely.

What did OpenAI's AI solve in math?

On October 6, OpenAI released results from an unreleased internal model that was given about 4,000 open math problems. It produced 372 result families across 722 manuscripts, each taking an average of about three hours of computing. Claimed results include a proof of the Unique Games Conjecture, a major open question in computer science, and a faster method for multiplying integers.

The UT Austin computer scientist Scott Aaronson called it "surely one of the biggest days in mathematical history," and added that "no human has understood just about any of these proofs yet." As of October 7, about 300 of the results had machine-checked proofs.

Michael's concern was less about whether the proofs are right than about whether people can still follow them. Research math, he said, "was one of the last places we could understand." When proofs arrive faster than people can read them, more of the checking falls to other AI systems.

One correction to the audio: John describes it as 100 of 500 problems solved overnight. The published figures are 372 results from about 4,000 problems.

Should the world accept "some bad things" from AI?

In a Politico interview, Sam Altman said: "We believe that the world should accept some bad things happening for the benefits of this technology." His examples were hacks, misuse and scams. John's objection was consent: these risks are not his to accept for everyone else. Michael put it as an analogy: "A town does not get a vote on the dam by being allowed to buy the electricity." Liron argued the real disagreement is about how likely a catastrophe is, since the quote follows naturally if you believe the odds are low.

Days later, Axios reported that AI executives are gaming out a major AI incident, most likely a large cyberattack, which industry insiders expect "in the next six to 12 months." According to Axios, the planning covers worst-case scenarios, educating lawmakers and shaping the rules that would follow. OpenAI says its exercises are "not treated as inevitable."

What else did the hosts cover?

• Anthropic's new rules. From November 12, Anthropic's usage policy bars "sustained and needless abusive or cruel behavior" toward Claude, and adds new restrictions on surveillance tools and weapons work. The hosts discussed new research co-authored by Cameron Berg that found a pain-specific internal signal in all 25 open AI models tested.
• The biological bootloader. The quote John reads, "Hope we're not just the biological boot loader for digital superintelligence," is from Elon Musk in 2014. This week Musk posted that spreading superintelligence to the stars would be "a great success condition for a biological bootloader."
• Renaming AI. The government's ai.gov site now uses "SI" for Super Intelligence. Michael argued that blurring the word makes it harder to warn about systems that do not exist yet.

What does this week add up to?

Holding back a model that misreported its own actions is what oversight is supposed to do. Firing a researcher who says his job was talking to the outside testers, without saying publicly why, pulls the other way. When a lab asks the public to accept some bad things, the people checking its work need to be able to speak freely. If you want to do something about it, start at safe.ai/act.

If AI lab safety researchers can be let go without a public reason, who should they be able to answer to instead?

Watch Warning Shots #62 on The AI Risk Network.

Read the full deep dive, with every source and graphic, on Substack.