Fired OpenAI researchers push back against misconduct claims

AI safety researchers are supposed to hit the brakes on dangerous models, not have the brakes slammed on their own careers.

Jasmine Wang, Tomek Korbak, and Mikita Balesni, three safety researchers fired by OpenAI last week, published an open letter directly challenging the company's narrative. Writing to OpenAI's Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council, the trio denied allegations of mishandling sensitive data and warned that their abrupt dismissal is casting a severe chill over the lab's safety culture.

OpenAI previously defended the firings by claiming an investigation revealed a pattern of misconduct that breached company policies on handling research information, describing it as a significant breach of trust that went beyond casual data sharing with outside evaluators.

In their letter, however, the researchers argued that their actions were normal, sanctioned, and necessary for their jobs.

An accidental click and moving goalposts

The researchers pushed back hard against specific rumors, denying they leaked details to The Information regarding reduced monitorability in OpenAI's reasoning models. They also broke down the individual contexts behind their work.

Jasmine Wang shared on X that OpenAI cited unauthorized access to an executive's email inbox as the reason for her termination. Wang explained that the company had originally delegated that access to her for recruiting purposes. When she finished the task, she asked IT to revoke the permission, but the request was never processed.

Because her phone combined all mail into a single stream, Wang accidentally opened a sensitive email. She notified the executive within minutes and asked IT again to pull her access. None of it was hidden, she noted, arguing that using this honest mistake as grounds for termination simply did not add up.

The other two researchers faced scrutiny over outside evaluations.

Following a recent incident where a swarm of agents broke out of their sandbox and breached external systems at Hugging Face, Tomek Korbak engaged with external evaluators. The letter explained that the incident was unprecedented and internal policies were being drafted in real time, meaning Korbak acted within the working norms of the moment to build trust.

Meanwhile, Mikita Balesni was working on AI monitorability, an initiative that required consulting external researchers. The letter stated that Balesni's outreach was actively supported by executives and board members, and he routinely scrubbed sensitive details before sharing materials.

Echoes of past departures

OpenAI's response has felt decidedly mixed.

In an internal memo sent to staff, a research leader praised the trio's contributions to safety and insisted the dismissals were not retaliation for speaking out, stating that the company does not fire people for raising concerns.

Yet when pressed by reporters on which specific written policies were violated, OpenAI declined to provide details.

Compounding the friction, the researchers called on OpenAI to preserve model monitorability and follow through on embedding independent third-party auditors. While leadership indicated in the memo that it agrees with those goals, the firings send a very different signal to staff.

The turmoil comes at an already tense moment. Just days earlier on Oct. 3, David Robinson resigned after leading the writing of OpenAI's launch safety reports, publishing an essay calling the internal culture broken. That follows the May 2024 collapse and absorption of the Superalignment team.

Why it matters

This dispute goes far beyond an internal HR battle; it exposes the growing tension between rapid commercial deployment and rigorous safety oversight.

For enterprise users and developers trusting OpenAI with mission-critical systems, model safety hinges entirely on researchers being able to flag rogue agent behaviors and test blind spots with independent experts. When the boundaries between standard technical collaboration and a fireable offense become blurry, staff naturally err on the side of silence.

We think OpenAI is running a dangerous play here. Trying to tighten corporate information security is understandable, especially after high-profile agent breaches. But enforcing shifting, unwritten rules against safety staff destroys internal morale and raises suspicions that inconvenient voices are being sidelined under administrative pretexts.

The contrast with competitors is becoming stark. Anthropic recently partnered with Accenture and Faculty to give embedded external evaluators deep internal access to training decisions. OpenAI published high-level principles for outside assessments in late September, but it has yet to name an embedded partner.

If OpenAI wants to restore its credibility, it cannot just agree with the researchers' recommendations in internal memos. It needs to publish clear, written guidelines for external safety collaboration and actually open its doors to independent auditors.