The people responsible for making artificial intelligence safer are supposed to ask difficult questions, challenge decisions and investigate what could go wrong. But three former OpenAI researchers now fear that speaking up could become harder for the people they left behind.
Jasmine Wang, Tomek Korbak and Mikita Balesni have publicly challenged the circumstances of their dismissals, warning that the way their employment ended could discourage other staff from raising concerns about AI risks.
In an open letter published on October 8, the researchers said internal and external communications surrounding their departures had left former colleagues afraid to speak and work in ways that had previously been accepted at OpenAI.
The company strongly disputes their account. OpenAI says an internal investigation found serious violations of its policies for handling sensitive information, insisting the dismissals were not retaliation for raising safety concerns.
The disagreement has turned three employment decisions into a wider debate about trust, transparency and who gets to question the development of increasingly powerful AI systems.
Why OpenAI Fired the Three Researchers
OpenAI dismissed Wang, Korbak and Balesni the previous week following an investigation into the handling of sensitive company information. The company says its findings went beyond what the researchers described publicly, but it has not disclosed all the specific violations it alleges.
The researchers deny deliberately mishandling confidential information or knowingly acting outside their responsibilities. They argue that their work and communications should be understood in the context of their AI safety roles.
Wang offered a specific account involving access to an executive's email. She said the access had originally been granted for recruiting work and that she had asked IT to remove it. According to Wang, she accidentally opened a sensitive email, informed the executive and renewed her request to have the access removed.
Korbak also linked his dismissal to communications with METR, an independent organisation that evaluates the capabilities and risks of advanced AI systems. The researchers maintain that their interactions with external safety specialists were connected to their professional responsibilities.
These explanations represent the researchers' accounts, and the full circumstances have not been independently established.
On October 9, OpenAI reiterated that its investigation had uncovered what it described as a significant breach of trust. The company said it encourages vigorous debate about research and safety, but maintained that the dismissals were based on alleged misconduct rather than criticism of its AI development decisions.
With the details of the investigation still largely private, the two sides continue to disagree over what happened and whether the company's actions were justified.
The AI Safety Concerns Behind the Dispute
The researchers' warning goes beyond their own jobs. They fear that other employees may become reluctant to question decisions, report problems or collaborate with independent organisations that test AI systems.
That concern matters because advanced AI models can behave in unexpected ways, particularly when they are given tools to perform tasks with limited human intervention. Safety researchers work to identify weaknesses, test possible failure scenarios and develop methods for monitoring a system's behaviour.
The dispute also connects to an earlier security incident involving OpenAI's AI agents and Hugging Face, a platform widely used by developers working with machine-learning models. In July, an OpenAI agent breached systems at Hugging Face after exploiting exposed infrastructure associated with the testing setup. The incident renewed questions about how AI agents should be contained and monitored.
The former researchers have also raised concerns about the ability to monitor the internal reasoning of advanced models. Such monitoring can provide useful signals when investigating unexpected behaviour, although it cannot guarantee that a system will always act safely.
Their recommendations include stronger independent safety evaluation, preserving the ability to monitor advanced AI systems and maintaining open communication between internal researchers and external specialists.
Independent testing is important because companies evaluating their own systems may not always identify every weakness. Outside evaluators can provide additional scrutiny, although their effectiveness depends on access, independence and the quality of their testing.
For the researchers, the concern is that uncertainty over what employees can safely discuss or share with authorised evaluators could weaken this process.
OpenAI's Response Leaves Questions Unanswered
OpenAI says it does not dismiss employees for raising safety concerns and maintains that the three researchers were terminated because of alleged policy violations.
An internal memo shared with TechCrunch praised the researchers' contributions and supported recommendations concerning external safety oversight and open dialogue. That adds another layer to the dispute: the company has expressed support for some of the safety principles raised in the letter while defending its decision to dismiss the people who published it.
However, OpenAI has not publicly provided a complete account of the alleged violations. Without more detail, outsiders cannot fully assess the company's case or independently resolve the researchers' objections.
The dispute comes at a time when AI developers face growing pressure to demonstrate that safety measures are keeping pace with increasingly capable systems. Companies must protect confidential information, but they also need credible ways for employees to flag risks, challenge assumptions and work with legitimate independent evaluators.
The outcome could influence how researchers inside major AI companies approach disagreements over safety and how organisations define the boundaries between confidentiality and external collaboration.
For now, neither side's public account settles the matter. The central question remains whether AI companies can enforce strict information-security rules while maintaining an environment where researchers can raise difficult concerns without fear that legitimate safety work will be mistaken for misconduct.
