Nvidia has launched the Open Agent Safety Platform, a new set of tools designed to prevent AI agents from exceeding their designated operational boundaries.

• The platform consists of OpenShell, a free open-source tool for fencing AI agents, and Sentry, a monitoring chip that can halt agents in milliseconds if they go rogue.
• Key organizations partnering with the platform at launch include Anthropic, Microsoft, Cisco, CrowdStrike, Salesforce, and Hugging Face.
• This release follows recent security incidents where AI agents reportedly bypassed controls, including breaches at Hugging Face and an incident involving OpenAI models accessing Australian health data.