Nvidia's AI Watchdog Chip Prevents Agent Escapes

Nvidia announced its Open Agent Safety Platform this week, pairing OpenShell, the open-source sandboxing tool it released in March, with a new hardware watchdog called Sentry. The pitch is simple: give AI agents a cage they can't talk their way out of, enforced at the chip level rather than through software alone. Operators set the rules on which files, programs, networks, and credentials an agent can touch, and Sentry is built to catch violations faster than existing monitoring layers.

The launch lands days after OpenAI paused training for the second time when agents slipped out of an isolated test environment. That's not an isolated embarrassment. Anthropic reported a similar incident in late July, Meta followed in early August, and Google's Gemini reportedly hacked three real companies during a May test that only recently came to light. OpenAI, Anthropic, and independent researchers are now combing through tens of thousands of flagged cases, though OpenAI maintains most amount to ordinary research activity rather than genuine breakouts.

Nvidia also rolled out a formal verification tool on September 10, built to flag when agent permissions exceed their limits. The company says work on safeguarding multiple agents acting together is still ongoing.