AI agent credentials live in the same box as untrusted code. Two new architectures show where the blast radius actually stops.
At RSAC 2026, cybersecurity leaders emphasized the need for enhanced governance and security measures for AI agents, advocating for a shift from traditional zero trust models to continuous action verification. Two companies, Anthropic and Nvidia, presented contrasting architectural approaches to secure AI agents, highlighting the critical importance of credential isolation and monitoring to mitigate risks associated with prompt injection and unauthorized access.
The most actionable insight for you is the critical need for AI agent architectures to adopt zero-trust principles, particularly in credential management and governance. Both Anthropic's Managed Agents and Nvidia's NemoClaw present different approaches to achieving this, with Anthropic focusing on structural credential isolation and Nvidia on robust monitoring and control. For your AI deployment strategies, prioritize architectures that structurally isolate credentials from the execution environment to minimize risk, and ensure your infrastructure can support the necessary observability and governance models to address indirect prompt injection vulnerabilities.