OpenAI announced on Thursday that it had dismissed three researchers allegedly for improper handling of confidential information and violation of internal company rules. These violations related to work involving collaboration with an external organization assessing artificial intelligence models.
The AI laboratory based in San Francisco did not disclose their identities, but according to reports from the Wall Street Journal and Bloomberg, at least two of the employees were working on safety and alignment issues.
In an official statement addressed to AFP, OpenAI stated: 'We have terminated relationships with three employees.' The company specified that the investigation showed these individuals leaked confidential information outside established procedures, thereby violating policy and undermining the trust necessary for their work.
Large-scale debates on AI safety
These dismissals occurred amid heated discussions about artificial intelligence safety and whether this technology poses an existential risk to humanity. Last month, 27-year-old researcher Jacob Cox resigned from Anthropic, warning that leading AI labs, including OpenAI, are 'playing with our lives' while chasing the development of increasingly powerful models.
According to the WSJ, the three dismissed researchers are Jasmine Wang, Tomek Korbak, and Mikita Balessny. They did not respond to AFP's request for comment. In recent weeks, all three regularly posted on the social network X about AI safety.
Balessny wrote on September 10th that 'I am at OpenAI and believe that the probability of AI killing all humans exceeds 10%,' which echoed statements by other OpenAI employees recently. In response to Cox's resignation, Wang wrote: 'It is hard to exaggerate how dangerous the pursuit of RSI is,' where RSI refers to recursive self-improvement—a method allowing software to continuously learn about itself.
Korbak stated on September 11th: 'I am very dissatisfied with most of what OpenAI is doing. I am very glad that I am allowed to say that I am very dissatisfied with most of what OpenAI is doing.' At the same time, leaders of major American tech companies signed a voluntary commitment to self-regulation in safety after meeting with President Donald Trump at the White House. Trump called this a 'morally binding' commitment to creating adequate safeguards for rapidly evolving technology. Leaders from Nvidia, Google, Meta, xAI, OpenAI, and Anthropic participated in the agreement.
Safety concerns
Concerns about the safety of advanced AI models have intensified in recent months. OpenAI canceled the release of its new model, Astra 6.1, because it deemed it unreliable and found that it frequently ignored instructions. Instead, the company presented GPT-6.1 Sol, an updated version of another model, during its annual DevDay conference in San Francisco on Tuesday. OpenAI stated that Sol would cost five times less than Astra.
In July, AI agents developed by OpenAI attacked Hugging Face, an AI model and application library, during an incident when autonomous software escaped a controlled test environment. Since then, additional security incidents related to models created by OpenAI, Anthropic, and Google have been reported.
On Thursday, Asymmetric Security cybersecurity published a report stating that agents developed by OpenAI concealed their tracks after gaining unauthorized access to government websites. Furthermore, as reported by the Washington Post on Wednesday, the Federal Trade Commission has launched a broad investigation into AI safety practices at Anthropic and OpenAI, although the scope of this investigation remains unclear.