OpenAI Fires Three Safety Researchers: Who Guards the AI Guardians?

OpenAI fired three safety researchers who say they were punished for warnings. The company denies it. Independent oversight is needed.

Three people whose job was to make artificial intelligence safer have lost their jobs. OpenAI, the maker of ChatGPT, says it fired Jasmine Wang, Tomek Korbak and Mikita Balesni because they broke its rules on handling sensitive information. The three say something different. They say they were fired for putting safety ahead of the company's short-term interests, and that their dismissal has made other staff afraid to speak up.

Both sides cannot be fully right. And the public cannot easily tell which one is. That is the real story. A handful of private companies are building some of the most powerful technology in the world. When their own safety staff raise alarms and then lose their jobs, the rest of us are left with a one-line statement from the company and a letter from the people who were let go. We deserve better than that.

What happened

The three researchers were dismissed late last week,. On Thursday, they published an open letter addressed to OpenAI's board and its safety committees. On Friday, OpenAI answered with a statement on social media.

The company said it had "parted ways" with the three after a "thorough investigation" found they "violated clear policies on handling sensitive information." It called it a "significant breach of trust," and said the problem went beyond what the researchers had laid out in their letter. It also said the firings were "not about them raising safety concerns," and that it agrees with the spirit of what the researchers wrote about preserving the ability to monitor advanced AI models.

The researchers tell a different story. Mikita Balesni wrote that he believes the three were "fired for prioritizing safety over the near-term interests of OpenAI as a corporation." He said his work included cross-company efforts on OpenAI's commitments to keep AI systems monitorable, and that he had taken care to remove sensitive details from materials before sharing them. Tomek Korbak has said he raised worries about a falling ability to monitor AI agents, the tools used to spot harmful behaviour. He said that on his exit call he was told the company no longer trusted him because he spoke too often with outside safety groups, which he took as a hint that he had leaked company secrets.

What the researchers are asking for

Their letter does not ask for revenge. It asks for three changes:

• Third-party auditors inside the company. Independent safety experts should be embedded within OpenAI, with real access.

• Keep AI monitorable. The company should preserve the ability to watch and understand what advanced models are doing.

• An open culture. Safety researchers should be able to speak with the wider safety community without fear.

These are modest asks. They are the kind of thing a company serious about safety would welcome. OpenAI says it agrees with the idea behind the second point. The question is whether it will accept the first and the third in practice.

Why this matters beyond one company

It would be easy to treat this as an office dispute. But this is not any company, and these are not any workers.

OpenAI is among the most valuable and influential technology firms., it is preparing for a stock market listing expected in 2027. Its yearly revenue run rate was close to $50 billion in September, below the roughly $70 billion it had suggested to investors earlier. A company in that position is under heavy pressure to grow, to launch products quickly and to please investors.

Safety work, by its nature, sometimes says "slow down" or "not yet." It is a cost, not a source of revenue. That is exactly why safety staff need protection. If they can be dismissed for talking with outside experts, then the people best placed to spot trouble will learn to stay quiet.

When a company marks its own homework on safety, the public needs someone else to check the answers.

The company's side deserves a hearing

Fairness requires a clear look at OpenAI's case. Companies have a right to protect real trade secrets, and the information that advanced AI labs hold can be sensitive. Leaking details of a model's capabilities or weaknesses could be harmful, or could help a rival, or in some cases help bad actors. A firm that fires someone for truly mishandling such information is not necessarily doing anything wrong.

OpenAI has also said that safety debates happen every day at the company, often spirited and critical, and that it cannot do its work without a high degree of trust. It has said it values the researchers' contributions and is "deeply sad" about the outcome.

All of that may be true. But it does not answer the central problem. We are asked to take the company's word that an internal investigation was fair. The investigation's findings have not been shared. The researchers deny the charge. In a dispute like this, "trust us" is not enough.

Why internal investigations are not enough

Internal investigations have a weakness. The company is the investigator, the judge and the party with the most to lose. Even if everyone acts in good faith, the process can look self-serving.

This is a familiar problem. In other fields, such as aviation, medicine and finance, we do not rely only on a company's own review when something goes wrong. There are outside regulators, independent investigators and legal protection for people who report problems. AI is now powerful enough that it needs the same.

The case also fits a pattern. Over the years, there have been several clashes between AI companies and their safety staff, including departures, public warnings and disputes over agreements that limited what former employees could say. Each time, the company has explained and moved on. But the pattern itself deserves attention.

The wider policy gap

The timing is awkward for the industry. Only days ago, AI executives met at the White House and agreed to a voluntary pledge to police themselves. President Donald Trump has dismissed AI safety fears as a "hoax." In California, the governor signed laws on worker protections and chatbot risk checks. As we understand it, California's law on frontier AI, signed in 2025, includes protections for employees who report serious risks. Whether those protections apply to this case will depend on facts and on the law, and the courts or regulators may have to decide.

The wider point stands. At present, in most places, an AI worker who sees a danger has few clear rights. There is no standard way to raise a concern outside the company, no guarantee of protection and no independent body that is built for the job.

India's interest

This might look like a faraway story, but India has a large stake. Indian companies, government departments and millions of users rely on AI products built by a few foreign firms. Indian engineers and researchers work for these firms, in India and abroad. Many Indian start-ups build on top of their models.

If the systems have hidden risks, such as unsafe behaviour, weak controls or poor monitoring, Indian users will feel the effects without having had any say. And if the people who spot such risks are not protected, India has no way to know.

India has general data protection and information technology laws, but, as far as we are aware, no specific law on AI safety or protection for AI whistleblowers. The government should look at this gap before the technology spreads further into public services, schools and workplaces.

What accountability looks like

The goal is not to punish a company for firing someone. It is to make sure that people who warn about risks are protected, that their claims are checked fairly and that the public can trust the answer. Here is what that would take.

1. Independent review of this case. An outside expert, agreed by both sides, should examine the dismissals and publish findings, with trade secrets protected.

2. Share the facts. OpenAI should explain, as far as it can without revealing real secrets, what rules were broken and how the investigation was done.

3. Accept third-party auditors. Give independent safety experts real access to models and internal safety processes, with reports to the public.

4. Protect those who speak up. Companies should not use confidentiality rules to stop employees from sharing safety concerns with regulators, auditors or independent safety groups.

5. Law for whistleblowers. Governments should pass clear laws that protect AI workers who report serious risks, and set up a safe channel for reporting.

6. A public record. Frontier AI firms should publish regular safety reports, including the number of internal safety concerns raised and how they were handled.

7. Keep models monitorable. Regulators should require that advanced systems stay open to inspection, so that harmful behaviour can be detected.

8. India's role. The Indian government should start work on rules for AI safety reporting and whistleblower protection, in consultation with researchers and industry.

The bottom line

Maybe OpenAI is right that it fired three people for a real breach. Maybe the researchers are right that they were let go for asking uncomfortable questions. The public cannot know from here. And that is the problem.

As AI becomes more powerful, the people who watch it for danger must be able to speak freely and safely. If they cannot, society loses its early warning system. A technology this important should not depend on the goodwill of a few companies, or on who has the better statement on social media. It should have rules, referees and a way for the truth to come out.