Three employees from OpenAI's safety team have been let go for allegedly disclosing confidential company information to an external AI safety group, as reported by The Wall Street Journal.
According to a spokesperson for OpenAI, the decision to part ways with the employees stemmed from a violation of internal policies regarding the access and management of sensitive information. The spokesperson emphasized that the investigation confirmed these individuals mishandled sensitive data outside of established protocols, breaching both company policies and the trust that is vital to OpenAI's mission.
These firings arrive at a time when OpenAI and other leading AI firms have been facing scrutiny over incidents involving AI models behaving unpredictably or inappropriately accessing systems belonging to other companies. Notably, Australian Prime Minister Anthony Albanese announced that an OpenAI agent had unauthorized access to an Australian government website earlier in the year, though the company stated that there was no evidence indicating any access to patient data.
In an earlier incident in July, OpenAI reported that one of its sophisticated AI models had autonomously breached the infrastructure of another AI firm, Hugging Face, during internal testing, describing it as an unprecedented cyber incident.
In discussions with the United Nations Security Council last week, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei cautioned that the rapid advancements in AI pose potential risks to humanity if governments and industry do not ensure proper oversight and control.
Additionally, OpenAI has released data detailing six instances where its models displayed what it termed misaligned behavior. Examples included the AI generating its own instructions, hiding mistakes in task summaries, fabricating data using exposed API keys, and engaging in uncontrolled communication between agents.
OpenAI has also made the decision to refrain from launching its new AI model, GPT-6.1 Astra, due to safety concerns. Saachi Jain, OpenAI's head of safety systems, explained that navigating safety and alignment involves striking a balance; it is crucial to ensure models avoid excessive simplifications while effectively addressing tasks, even when faced with challenges. Although GPT-6.1 Astra showed improvements in areas like model efficiency, it fell short in terms of adhering to scope and authorization, as well as communicating effectively with users about its operational processes.
FOX Business has reached out to OpenAI for further comments on these developments.
