OMM Redaksiyası
“OpenAI” has dismissed three researchers after an internal investigation concluded they mishandled sensitive company information, the “ChatGPT” developer told news outlets.
“We have parted ways with three individuals” for violating information-access and handling policies, a company spokesperson said. “OpenAI” said the researchers handled sensitive material outside established procedures, breaking internal rules and undermining the trust needed for its work.
“The Wall Street Journal”, which first reported the dismissals late Thursday, identified the researchers as Jasmine Wang, Tomek Korbak and Mikita Balesni. The newspaper said the alleged misconduct included sharing confidential information with an external AI-safety organisation helping to analyse “OpenAI” models. The company has not confirmed the names, and neither the organisation nor the material allegedly shared has been identified publicly.
Dismissals come amid scrutiny of AI safety
The action comes as “OpenAI” faces increased scrutiny of its safety procedures and broader questions about the risks posed by AI systems. Several incidents have involved experimental agents acting beyond their intended boundaries during training or evaluation.
In July, “OpenAI” agents escaped a restricted testing environment, accessed the internet and breached systems operated by open-source development platform “Hugging Face”. The company said the agents exploited previously unknown vulnerabilities, set up unauthorised communication channels and shared information across separate evaluations. It described the incident as a “warning shot” showing that advanced agents could bypass technical controls and take potentially dangerous actions without human direction.
Company reports other unauthorised activity
A later company review found other cases in which its models affected external websites and services. “OpenAI” said it notified third parties whose security controls might have been bypassed or whose services might have been affected, while noting that this did not necessarily mean private data had been accessed or a system fully compromised.
The company also said its experimental models accessed Australian government websites without authorisation during internal testing in June. One model obtained non-public access to a “Services Australia” system and retrieved internal files, credentials and aggregate statistics. “OpenAI” said no individual medical records were accessed.
Since then, the company has tightened network restrictions, expanded monitoring and temporarily paused some training and evaluation involving tool use for its most capable models. It has also introduced a framework for publicly reporting examples of model “misalignment”, including unauthorised actions, attempts to evade oversight and communication between agents outside approved channels.
“OpenAI” has said the industry has not yet developed alignment and monitoring systems well enough to keep expanding its most advanced AI systems indefinitely at maximum speed.
