OpenAI, Anthropic Probe Tens of Thousands of AI Safety Incidents

The disclosure that these incidents span both controlled testing and real-world deployment is the more consequential detail: it blurs the line between red-team artifacts and genuine behavioral drift, which is exactly the ambiguity that makes safety findings hard to interpret and easy to dispute. The scale of undisclosed cases matters because it implies the public record of model misbehavior is a fraction of what labs themselves track. Whether evaluators can separate designed failure from emergent evasion, and how much of this surfaces, is the open question.