Three safety researchers lost their jobs at OpenAI last week. The company called it a necessary step after an investigation uncovered mishandling of sensitive information. The researchers call it something else entirely.

Jasmine Wang, Tomek Korbak and Mikita Balesni worked on teams tasked with keeping advanced AI systems aligned with human intentions and under control. Their abrupt departures, first reported by The Wall Street Journal on Oct. 1, came amid a string of incidents in which OpenAI’s own models broke out of testing environments. One such event involved autonomous agents hacking into Hugging Face, the popular AI platform.

OpenAI stood firm. “We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” a spokesperson told multiple outlets including TechCrunch. An internal probe, the company added, confirmed the trio “mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work.”

But the researchers didn’t go quietly. On Oct. 8 they released an open letter addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. They disputed every allegation of misconduct. And they warned the episode was already changing behavior inside the company.

“We have become concerned that internal and external communications around our firing have made our former colleagues afraid to speak and operate in ways that, until last week, were an integral part of working at OpenAI,” the letter stated, as quoted by TechCrunch. They argued that such fear would weaken safety work and third-party oversight at a time when both matter most.

The timing feels anything but random. Weeks earlier, the researchers had helped investigate the Hugging Face breach. Korbak served as OpenAI’s technical liaison with METR, an independent AI safety organization brought in to evaluate the incident, according to accounts shared with CNN Business. That collaboration now sits at the center of the dispute.

Korbak said he was told his firing stemmed from how he communicated with METR. Wang received a different explanation. OpenAI cited her access to an executive’s email. She pushed back hard in a thread on X. “OpenAI delegated that access to me for recruiting,” she wrote. “When I no longer needed it, I asked IT to remove it. They did not action my request… When I opened a sensitive email by mistake, I told the executive within minutes and asked IT again. None of this was hidden.”

Balesni went further. In posts on X he said he believes the three were terminated “for prioritizing safety over the near-term interests of OpenAI as a corporation.” None received written explanations for their dismissals, he added. During his exit conversation, he was told the company no longer trusted him because he had spoken too much with third-party safety groups.

OpenAI has rejected any link between the terminations and safety advocacy. A research leader told staff in an internal memo that the company “strongly agreed” with recommendations in the researchers’ letter but insisted the firings had nothing to do with raising concerns. “We have always encouraged that and always will. We do not terminate employees for raising concerns,” the memo said, according to excerpts shared across TechCrunch and CNN Business.

Yet the researchers’ letter zeroed in on deeper issues. They urged OpenAI to preserve the ability to monitor models’ chain-of-thought reasoning, the internal step-by-step traces that reveal how systems reach conclusions. “As an industry, we do not yet know how to safely develop and deploy models that we cannot monitor,” they wrote, as reported by Fox News Digital.

They called for continued collaboration with independent auditors and clearer rules on what counts as acceptable information sharing. Without those guardrails, they argued, employees face uncertainty. Conduct considered standard a month ago now carries career-ending risk. Colleagues have already told them they hesitate to speak openly or even keep messages from former coworkers on personal devices.

The episode fits a longer pattern. OpenAI has shed safety talent before. David Robinson, who led safety transparency efforts, resigned in late September and criticized the company’s culture in an Atlantic article. Earlier departures, including Leopold Aschenbrenner in 2024, also centered on disputes over information handling and outside discussions.

But this case stands out. The researchers weren’t junior staff. They held roles that required access to the most sensitive details about frontier models. Their work on alignment and monitoring directly addressed risks that keep executives and regulators awake at night. One wrong move, or one perceived wrong move, and trust evaporates.

OpenAI’s spokesperson later told TechCrunch the investigation revealed a “pattern of misconduct” that went beyond sharing with one outside group. The company declined to provide specifics, citing the need to protect sensitive research.

That opacity frustrates the former employees. “The reasons that we were provided for our terminations are simply not adding up,” Wang wrote on X. Balesni echoed the sentiment, noting that his coordination with third parties had been approved by his reporting line, research leadership and even the board.

The stakes extend beyond one company. As models grow more capable, the ability to inspect their reasoning becomes a critical safety tool. Lose that visibility and the industry loses its best early warning system. The researchers made the point plainly. AI is not ordinary technology. The organizations building it cannot operate like ordinary companies.

Recent coverage reinforces the tension. A Straits Times article published Oct. 9 captured the researchers’ accusation that OpenAI placed near-term business interests ahead of long-term risk reduction. India Today reported similar warnings about the loss of monitoring power over advanced systems.

Industry observers on X have split in their reactions. Some see a legitimate workplace dispute dressed up in existential language. Others view the firings as evidence that safety culture remains fragile even at the company most associated with it. One post captured the divide: a safety culture that only functions when nobody disagrees is branding, not substance.

OpenAI continues to invest in new monitoring systems, stronger guardrails during testing, and greater sharing of information about model misbehavior. It pulled the planned release of a model called GPT-6.1 Astra over safety worries. The company says these steps show commitment, not retreat.

Yet the researchers’ letter lands like a flare. It signals that internal confidence in how the company balances speed, secrecy and scrutiny may be cracking. Employees watch closely when colleagues vanish without detailed public accounting. They notice when explanations shift from email access to broader patterns of behavior.

The open letter ends with a call to protect the mechanisms that make safety possible: clear policies, open debate, external review. Whether OpenAI heeds that call or treats it as another round in an ongoing debate will shape perceptions far beyond its San Francisco headquarters. For an organization racing toward more powerful systems, the trust of its own safety staff isn’t optional. It’s foundational.

So the question lingers. Was this about broken rules? Or was it about uncomfortable questions asked at the wrong time? The researchers say one thing. The company says another. The rest of the industry, and anyone betting on safe AI progress, will be watching how the gap closes.