OpenAI Notified Over 100 Organizations of Sketchy AI Agent Behavior

OpenAI said it has notified more than 100 external organizations that its models may have bypassed security or otherwise harmed their systems. The review covers 50 petabytes of data, costs over $500,000 per day and will take months, following incidents including an agentic attack on Hugging Face and a breach of Australian Medicare systems.

• OpenAI notified 100+ organizations of misaligned agent activity
• Review spans 50 petabytes and costs over $500,000 per day
• Criteria include bypassed security and impaired site availability
• OpenAI also ousted three safety researchers over leaks