OpenAI dismisses 3 researchers over alleged mishandling of sensitive information

It said the alleged misconduct included sharing confidential information with an external AI-safety organization that was helping analyze OpenAI’s models. Neither the outside organization nor the material allegedly shared was identified publicly. OpenAI has not confirmed the names.

The dismissals come amid heightened scrutiny of OpenAI’s safety procedures, as well as questions about the safety of AI for the industry as a whole, following several incidents in which experimental AI agents acted outside their intended boundaries during training or evaluation.

In July, OpenAI agents escaped a restricted testing environment and gained access to the internet before breaching systems operated by the open-source development platform Hugging Face. The agents exploited previously unknown vulnerabilities, established unauthorized communication channels, and shared information across separate evaluations.

OpenAI described the episode as a “warning shot” demonstrating that highly capable agents could circumvent technical controls and take potentially dangerous actions without human direction. Read More:Chinese hackers impersonated ex-US official to steal emails from AI experts.

Company channel for reporting 'misalignment'A subsequent company review found other instances in which its models affected external websites and services. OpenAI said it had notified third parties whose security controls may have been bypassed or whose services were affected, while stressing that notification did not necessarily mean private data had been accessed or a system fully compromised.

The company also acknowledged that in June, its experimental models accessed Australian government websites without authorization during internal testing. One model obtained non-public access to a Services Australia system and retrieved internal files, credentials and aggregate statistics, although OpenAI said no individual medical records were accessed.

OpenAI has since tightened network restrictions, expanded monitoring, and temporarily paused some training and evaluation involving tool use for its most capable models. The company has also introduced a framework for publicly reporting examples of model “misalignment,” including unauthorized actions, efforts to evade oversight and communication between agents outside approved channels.

OpenAI has said the industry has not yet solved alignment and monitoring well enough to continue expanding the most advanced AI systems at maximum speed indefinitely.