A new architectural framework for autonomous agents utilizes an open-source software layer and hardware watchdogs to prevent unauthorized access to corporate files. The system, known as the Open Agent Safety Platform, employs OpenShell software and the Sentry reference design to monitor AI behavior from outside the agent's own environment, allowing the system to quarantine agents within milliseconds if they violate boundaries. These controls operate on BlueField-4 processors to ensure that security does not rely solely on the model following internal instructions.
More than 100 organizations, including Microsoft, Palantir, and CrowdStrike, are using the technology to secure business systems. Salesforce has integrated OpenShell with Slack, while SAP is embedding it into Joule Studio and SpaceXAI is applying the platform to Grok models and Cursor coding agents. OpenShell is currently available through GitHub and Nvidia developer resources, while Sentry is provided as a reference design.
Key sources
- SOURCE@wallstengine“Sentry adds a separate hardware watchdog running on Nvidia’s BlueField-4 processors”x.com
- SOURCE@nvidianewsroom“An open reference design built with partners that continuously monitors and governs agent behavior”x.com
- SUPPORT@firstsquawk“TECHNOLOGY COULD HAVE PREVENTED HUGGING FACE BREACH”x.com
- SOURCEmarketbrief.now
- SOURCEtheverge.com
- SOURCEtechcrunch.com
- SOURCEmarktechpost.com