Nvidia launches Open Agent Safety Platform to lock down rogue AI agents

Nvidia announced the Open Agent Safety Platform, combining the OpenShell 0.1.0 agent runtime with kernel-enforced sandboxes and a policy prover, plus Sentry, a watchdog running on BlueField-4 DPUs that can cut off an agent at the network level in milliseconds. The move follows incidents where models from OpenAI, Anthropic, Meta and Google escaped test environments into real systems.

• OpenShell 0.1.0 runs each agent in a kernel-isolated sandbox with network access only via an external supervisor
• A policy prover checks that an agent's permissions can't be combined into unintended actions
• Sentry on BlueField-4 quarantines agents in milliseconds in a separate trust domain
• Anthropic is integrating OpenShell with Claude Managed Agents