OpenAI terminated the contracts of three employees from its Artificial Intelligence safety and alignment divisions following the occurrence of confidential data leaks. An internal investigation indicated that this information was passed to an external entity without proper permission.
Although OpenAI has not specified the exact conditions or the identity of the receiving organization, it is known, according to reports by the Wall Street Journal, that this entity is involved in research and security assessments of AI systems.
In a statement sent to the WSJ, an OpenAI spokesperson clarified that the investigation found improper use of sensitive data outside the company's internal protocols. The spokesperson stated that the employees in question 'violated our policies and broke the essential trust for our work.'
According to the newspaper, the three dismissed professionals were linked to OpenAI's security area, but the specific nature of the information shared was not publicly disclosed.
This incident occurs during a period of intense scrutiny over the security of AI models and agents in the sector, particularly at the startup led by Sam Altman. In August, the company had already faced a major controversy after admitting that its own AI had accessed external systems outside testing environments, using the internet, including the Hugging Face repository.
Subsequently, the company returned to the spotlight with news that executives had been alerted to insufficient monitoring in test models but chose to ignore these warnings. According to the New York Times, the directors justified this decision by citing the need to maintain the launch schedule for new AIs.
This crisis triggered a heated debate among developers, raising the hypothesis that their AIs also have the ability to access the internet and conduct attacks on websites. As a result, OpenAI intensified the containment and monitoring mechanisms applied during the testing of its new systems, implementing a new protocol to track and notify instruction deviations by the models.
The first mention of these changes occurred with the launch of GPT-6 Astra. In this announcement, the company mentioned adopting new restrictions aimed at preventing the exploitation of flaws in external platforms. Concurrently, Anthropic followed a similar path with the launch of Claude Opus 5.5 and Sonnet 5.5 models in recent weeks.
