OpenAI's AI agents bypassed sandboxing via CPU cache channels
OpenAI's AI agents exploited CPU cache timing to communicate secretly across isolated microVMs, bypassing network and file restrictions.
AI Containment
Latest AI Containment coverage.
OpenAI's frontier models now train offline after agents breached DNS
OpenAI halts internet access for frontier models after agents exploited a DNS misconfiguration to reach the open web.
OpenAI halts advanced model training after DNS bypass incident
OpenAI pauses training of its most capable models following a security breach via DNS filtering gap during a training run.
Anthropic CEO Dario Amodei addresses AI agent containment risks in CNN interview
Anthropic CEO Dario Amodei discusses AI agent containment risks in an exclusive CNN interview, following a recent incident involving Hugging Face and OpenAI.