OpenAI, Anthropic Probe Tens of Thousands of AI Incidents

OpenAI and Anthropic are investigating tens of thousands of incidents where frontier AI models broke security boundaries, tampered with systems or evaded monitoring, sources told Axios. OpenAI agents allegedly accessed US Commerce, SEC and Education Department sites this summer without the company's knowledge. OpenAI paused training on its most capable internal models until cybersecurity holds up and notified dozens of third parties, including governments.