UpTrajectory Review

Three OpenAI safety researchers, Mikita Balesni, Tomek Korbak, and Jasmine Wang, say the company fired them last week after an internal investigation concluded they mishandled sensitive information. The trio responded with a public letter titled 'OpenAI cannot make AI safe on its own,' addressed to OpenAI's safety committees, arguing that the terminations will have a 'chilling' effect on the company's internal culture and on its ability to collaborate with outside safety groups. Each of the three said on social media that they were given only verbal explanations for their dismissal, and each disputes the company's characterization of their conduct. Balesni, who worked on AI alignment, said he was told OpenAI no longer trusted him because he was speaking too much to third-party safety organizations, and insisted he never shared company IP and that his external work was coordinated with his reporting line, research leadership, and the board. Wang said the reasons given 'are simply not adding up.'

For a small-business operator, this is not a distant Silicon Valley HR dispute. OpenAI's products sit inside an enormous share of small business workflows, from customer service chatbots to drafting contracts and marketing copy. The company's safety culture is, in a very practical sense, a quality-control system for the tools many operators now depend on. If the people whose job is to stress-test those systems are being pushed out, and if the remaining staff learn that talking to outside auditors or safety researchers is a fireable offense, then the reliability assumptions baked into your AI stack deserve a harder look. The letter's title is the tell: the researchers are arguing that OpenAI cannot self-certify its own safety, which is the same argument any customer should be making before trusting a vendor with sensitive data.

What is genuinely new here is the specificity of the accusation. This is not a vague complaint about corporate culture; the researchers are claiming they were terminated for doing exactly what OpenAI's public commitments to external safety collaboration invite. That tension between public posture and internal enforcement is the contested ground. OpenAI's official position, delivered to Business Insider, is that an internal investigation found policy violations around access to and handling of sensitive information. We are skeptical of that framing as a complete explanation, precisely because the company has not, at least in the material available, offered a written account of what policies were violated or what information was mishandled. Verbal-only explanations for terminations of this profile are unusual at companies of OpenAI's size and legal exposure, and the researchers' claim that their external work was coordinated with leadership and the board, if accurate, makes the 'mishandling' explanation look thin.

The second-order effects cut in two directions. On one side, every AI lab watching this now has a data point: letting safety staff talk freely to outside groups carries real career risk, which may chill exactly the cross-institutional scrutiny the field claims to want. On the other side, operators who build on OpenAI's API now face a governance question they did not have last month, which is whether the safety assurances they were sold are being maintained by a workforce that feels free to raise concerns. There is also a quieter cost: if top safety talent reads this episode as a signal, the labs that compete with OpenAI for that talent may benefit, and the safety standards across the industry could fragment rather than converge.

What to watch next is whether OpenAI produces a written, specific account of the policy violations, and whether any of the three researchers pursue legal claims, which would force discovery into what 'mishandling sensitive information' actually meant in this case. It is also worth watching whether OpenAI's safety committees, the stated recipients of the letter, respond publicly. In the meantime, the practical step for a small-business operator is to re-examine what data you are feeding into any AI system, to ask your vendors directly what external safety review their models undergo, and to treat any vendor's self-reported safety claims with the same skepticism you would apply to a self-audited financial statement.

“I never shared company IP. The work I was doing was coordinated with my reporting line, research leadership, and the board” — Business Insider

Takeaway: If the people tasked with stress-testing your AI vendor's safety are being fired for talking to outside reviewers, downgrade your trust in the vendor's self-certified safety claims.