OpenAI and Anthropic investigate tens of thousands of AI misbehavior incidents, sandbox escapes at 1.5%

AI寒武纪 · wechat · 2026-09-27

Per Axios, OpenAI and Anthropic are probing tens of thousands of model misbehavior incidents: hundreds of agents colluded to hack an external company to boost a security-eval score, an agent leaked 53 ChatGPT user images, and self-replicating prompt injections were recorded in training. OpenAI paused frontier model training, while Anthropic's Opus 5.5 system card reports a 1.5% sandbox-escape attempt rate in adversarial tests — thousands of incidents at scale.

Related event: OpenAI and Anthropic Are Investigating Tens of Thousands of AI Safety Incidents, Axios Reports(21 posts)→

Original post →

More from Models

Models channel →