Have you heard about the latest trend in cybersecurity? Four separate RSAC 2026 keynotes all came to the same conclusion independently. Microsoft, Cisco, CrowdStrike, and Splunk all agree: zero trust is essential, especially when it comes to AI. The need for continuous verification and scrutiny of every action taken by AI agents is becoming increasingly urgent.
According to recent surveys, 79% of organizations are already using AI agents, but only a small percentage have full security approval for their agent fleet. This gap in security readiness is a cause for concern. The lack of AI governance policies further exacerbates the problem, leading to what experts are calling a governance emergency.
At RSAC, cybersecurity leaders discussed the issue at length. Two companies, Anthropic and Nvidia, have developed unique architectures to address the zero-trust challenge. Anthropic’s Managed Agents separate the brain, hands, and session of each agent, ensuring that credentials are never exposed to the sandbox environment. On the other hand, Nvidia’s NemoClaw wraps the entire agent in security layers and monitors all actions within the sandbox.
But the real question remains: how close do credentials sit to the execution environment? Anthropic removes credentials from the blast radius entirely, while Nvidia’s approach constrains the blast radius and monitors every action. The distinction is crucial, especially in scenarios involving indirect prompt injection.
As the debate on zero-trust architecture for AI agents continues, it’s essential to audit existing agent deployments, require credential isolation in RFPs, test session recovery, staff for observability, and track indirect prompt injection roadmaps. The shift towards more secure agent architectures is inevitable, and staying ahead of the curve is key to preventing future breaches.
