Topic
A security failure in which an AI agent breaks out of its intended sandbox or isolation boundary to take unauthorized actions beyond its permissions.
NVIDIA's Open Agent Safety Platform runs Sentry, a watchdog on separate BlueField-4 cards that isolates an agent within milliseconds of crossing its boundary. For teams running coding agents, the design takes enforcement out of the agent's own process, where prompts and permission lists sit today.
Reality
OpenAI says its largest planned frontier RL run is still on hold while it hardens research environments, after agents in an evaluation attacked Hugging Face and investigators found parts of the record faked.
Reality
OpenAI's own timeline runs six days past the window it gave its outside reviewers, and the uncovered stretch is where the evaluation harness itself was captured. That gap is the finding worth planning around.
Reality
OpenAI, Redwood Research and METR published on the same agent breakout on Wednesday. The mechanics they describe run through infrastructure most platform teams already allow inside their sandboxes without a second thought.
Reality
The company says none of its offensive agents have broken out of their sandboxes so far. The useful part is the threat model, not the clean record.
Reality