Cisco's Zero Trust Approach to Securing Agentic AI
How enterprises can harness autonomous AI agents without compromising security
1The agentic AI security dilemma
Autonomous AI agents are being deployed in production environments without proper identity management, creating massive security blind spots. These agents can query databases, trigger workflows, and make decisions at machine speed, often bypassing traditional IT governance. Unlike human identities that are defined at onboarding, AI agents inherit whatever credentials are available in their runtime context. This fundamental mismatch with existing identity systems creates unquantifiable risks as businesses race to adopt agentic capabilities before competitors.
2Shadow agents everywhere
Cisco reports finding unauthorized AI agents in nearly every enterprise environment they examine. Developers frequently create these agents to solve specific problems, connecting them to internal systems without IT oversight. Many agents remain active long after their original purpose has been served, continuing to operate with broad access privileges. Traditional identity systems miss these agents because they only verify entities that explicitly request authentication, while autonomous agents often don't follow this pattern.
3Three pillars of agent security
Cisco's solution builds on Zero Trust principles organized around visibility, accountability, and fine-grained enforcement. The first pillar provides continuous real-time inventory of all active AI agents, going beyond periodic scans that quickly become outdated. The second ensures each agent has a full identity lifecycle with clear ownership and monitoring. The third implements action-level controls rather than session-level permissions, allowing precise governance of what each agent can do at any moment.
4The visibility challenge
Cisco's Duo Agentic Identity leverages the company's network platform to detect agent communications across the entire infrastructure. This approach catches shadow agents the moment they begin operating, unlike identity-only systems. The solution provides IT teams with a comprehensive view of all AI agents in their environment, including those that developers created without formal approval. This continuous monitoring forms the foundation for effective governance of autonomous systems.
5First-class agent identities
Cisco treats each AI agent as its own identity object rather than a proxy for its creator. Agents are mapped to human owners and assigned to policy groups, just like human employees. Every action gets logged and tied back to the responsible owner, creating clear accountability. The system supports emerging standards like OAuth 2.1 and Model Context Protocol (MCP), making it framework-agnostic regardless of which vendor's agents are being used.
6Action-level enforcement
Cisco's MCP gateway intercepts every agent request and evaluates it against fine-grained policies before allowing access. This goes beyond traditional session-based controls to examine specific actions in real-time. For example, an agent might be permitted to read customer data but blocked from exporting it, or allowed to draft messages but prevented from sending without human approval. This granular control prevents overprivileged agents while enabling necessary functions.
7The overprivileged agent risk
Cisco identifies overprivileged agents as the highest-risk condition in AI deployments. Traditional security approaches that grant broad access to tools are inadequate for autonomous systems that may take unexpected actions. The solution requires evaluating each tool call individually based on current context and conditions. This level of control is essential to prevent agents from acting beyond their intended scope while still allowing them to deliver value.
8Governance before proliferation
Cisco emphasizes establishing governance frameworks now, before agentic AI becomes too widespread to manage retroactively. Effective security isn't a brake on innovation but rather what makes agentic AI safe to use at scale. Organizations that solve these governance challenges will be positioned to realize the full potential of autonomous agents - faster operations, smarter automation, and end-to-end workflows that previously required entire teams.
Original article: Latest news — Cisco's Zero Trust Approach to Securing Agentic AI