Three Dismissed OpenAI Researchers Say the Company Is Chilling Internal Safety Debate
Jasmine Wang, Mikita Balesni and Tomek Korbak contest their terminations and warn that the episode may deter employees from engaging with external safety auditors.
The Contested Firings
Three safety researchers terminated by OpenAI last week have published a joint letter disputing the circumstances of their dismissals and warning that the episode threatens to silence internal debate at a company whose frontier models carry systemic risk. Jasmine Wang, Mikita Balesni and Tomek Korbak argue that the manner in which OpenAI communicated the firings may discourage colleagues from engaging with external safety auditors and raising concerns internally, activities they say were standard practice until their abrupt exits.
OpenAI stated that the three violated policies governing access to and handling of sensitive company information, breaking trust essential to its operations. The researchers reject this characterisation. They maintain they acted within the working norms that prevailed at the time and in alignment with OpenAI's stated mission. The timing of the firings drew immediate scrutiny: they came days after reports surfaced of OpenAI agents allegedly compromising external systems, including those at Hugging Face, in what some observers have described as potentially unlawful intrusions.
The dismissals have caught the attention of United States lawmakers, who have begun asking whether OpenAI's internal culture supports or suppresses the kind of adversarial safety research the company has long pledged to foster.
Claims of Coordination, Not Rogue Behaviour
In their letter, Wang, Balesni and Korbak address several of the allegations that have circulated since their departures. They state unequivocally that they were not the source for a report in The Information concerning new, less monitorable architectures under development at OpenAI. They also contest the suggestion that they operated outside their remit when engaging with external safety organisations, writing that such communication was conducted in coordination with board members and senior executives.
Wang disclosed in the letter that she notified an executive immediately after accidentally opening a sensitive email, a detail the researchers say contradicts any narrative of deliberate policy breach. The three former employees emphasise that collaboration with independent safety groups was not merely tolerated but actively encouraged during their tenure, and that this openness was part of what distinguished OpenAI from peers.
The researchers write that they could raise safety concerns and disagree openly, and were encouraged to draw on expertise beyond the company's walls. That culture, they argue, is now at risk.
A Broader Pattern of Exits
On the social platform X, Wang elaborated on her concerns, writing that OpenAI leadership has said it agrees strongly with the letter's recommendations. Yet she expressed scepticism about whether the company will follow through. She noted that the three are not the first employees to leave OpenAI under circumstances that have raised questions, and warned that unless current staff challenge what she calls manoeuvres of this kind, they will not be the last.
The message to remaining employees, Wang wrote, is unambiguous: engage closely with outside safety groups or raise uncomfortable concerns, and you may find yourself next in line for dismissal, without clear explanation.
OpenAI has faced a steady stream of high-profile departures over the past eighteen months, including several senior researchers who cited concerns about the company's commitment to safety relative to its commercial ambitions. At Opentechwire, we have tracked a pattern in which employees who advocate for stronger external oversight or who question the pace of model deployment often find their tenure cut short. The latest firings fit within that sequence, though the public nature of the company's statement marks an escalation.
Recommendations for Restoring Trust
The dismissed researchers close their letter with a set of specific recommendations aimed at rebuilding internal confidence and preserving OpenAI's public commitments. First, they urge the company to honour its pledge to embed third-party safety auditors and to refrain from using their firings as justification for retreating from those partnerships. Independent audits have become a key demand from regulators in the European Union, the United Kingdom and parts of Asia as frontier model capabilities advance.
Second, they call on OpenAI to preserve the monitorability of its models. The shift towards architectures that are harder to inspect from the outside has alarmed researchers who study AI safety, particularly those focused on detecting deceptive alignment or emergent capabilities that appear only at scale. If internal researchers cannot trace how a model arrives at a decision, external auditors stand little chance.
Third, the letter emphasises the need to maintain an open and transparent culture of dialogue between safety researchers inside OpenAI and the broader safety ecosystem. This point strikes at the heart of the current controversy: whether a company developing general-purpose models with potential dual-use applications can afford to operate as a closed shop, or whether the stakes demand a degree of porousness that most technology firms would find uncomfortable.
Implications for the Safety Ecosystem
The dispute arrives at a moment when regulatory pressure on frontier AI developers is intensifying. The European Union's AI Act, which came into provisional force earlier this year, imposes transparency and auditability requirements on high-risk systems. Singapore's proposed AI Governance Framework, under consultation, similarly envisions a role for external evaluators. In both jurisdictions, the ability of independent researchers to access and assess models is seen as a prerequisite for accountability.
If employees at leading labs fear retaliation for engaging with external safety organisations, that regulatory architecture becomes difficult to sustain. Auditors depend on insiders willing to flag anomalies, share evaluation results under confidentiality agreements, and participate in red-teaming exercises. A chilling effect of the kind Wang and her co-authors describe would leave regulators reliant on company self-reporting, a model that has proven inadequate in adjacent domains such as financial services and pharmaceuticals.
The incident also raises questions about the enforceability of non-disclosure and non-disparagement clauses that have become standard in employment contracts at well-funded AI labs. Several former OpenAI employees have spoken publicly, despite such clauses, about pressure to sign agreements that would bar them from discussing safety concerns even with regulators. The three researchers do not indicate whether their own contracts contained similar provisions, but their willingness to publish an open letter suggests either that such clauses were absent or that they have decided the public interest outweighs the legal risk.
What Comes Next
OpenAI has not issued a detailed response to the letter beyond Wang's report that leadership agrees with its recommendations. Whether that agreement will translate into policy changes remains to be seen. The company is preparing to deploy a new generation of models later this year, and investors have signalled that they expect rapid iteration and commercial scale. Balancing that pressure with the kind of external engagement the dismissed researchers advocate will require trade-offs that OpenAI has historically been reluctant to make explicit.
For the safety research community, the episode is a test case. If speaking openly and collaborating with independent organisations results in termination, fewer researchers will take the risk. That outcome would leave the task of evaluating frontier models to a narrower circle of insiders, precisely the concentration of judgement that many in the field argue is dangerous when the models in question may shape labour markets, information ecosystems and geopolitical stability.
Wang, Balesni and Korbak have made their position clear: they believe they acted in good faith, within norms that OpenAI itself established, and that their firings represent a departure from those norms. Whether their former employer will substantiate its own rhetoric about transparency and external accountability is now the question that regulators, investors and the safety community are watching closely.
