Cisco has unveiled a security framework encompassing identity management, access control, pre-validation, and operational monitoring to mitigate security issues hindering the adoption of AI agents in enterprises. The company determined that as AI moves beyond providing answers to performing actual tasks, controlling the authority and scope of agent actions has emerged as a critical challenge.

Cisco announced the related features at RSAC 2026, which opened in San Francisco, USA, on March 23 (local time). The company stated that in its survey, 85% of major enterprise customers are testing AI agent pilots, but only 5% have expanded to the actual production stage.

At the heart of the technology is the application of zero trust for AI agents. Cisco has proposed a method that links Duo IAM with Secure Access to assign an identity to each agent, connect them with a human representative, and grant granular permissions for short periods only for necessary resources. In addition, it was designed to track agent behavior and access paths more precisely through a structure that passes all tool traffic through the MCP gateway.

Pre-deployment verification capabilities have also been enhanced. 'AI Defense: Explorer Edition' is a self-service tool that enables developers to test models and applications using multi-turn-based attack scenarios to check for vulnerabilities such as prompt injection or jailbreaking. Along with this, the Agent Runtime SDK, an LLM security leaderboard for comparing model risks, and the open-source framework 'DefenseClaw' were also released. DefenseClaw is expected to integrate with NVIDIA OpenShell to support the automation of security checks and asset inventory tasks.

The scope of application extends beyond the AI development phase. Cisco has presented an "Agentic SOC" concept based on Splunk, which includes exposure analysis, Detection Studio, federated search, and various specialized agents. This configuration focuses on enabling security operations centers to connect distributed data to identify threats faster and automate certain response procedures. Some features have already been released, while the remainder are scheduled to be unveiled sequentially between April and June.

This announcement can be seen as an example of presenting security standards necessary for companies to move generative AI experiments into actual business systems, amidst the visible proliferation of AI agents. In particular, there is a growing possibility that the trend of clarifying the identity and accountability of agents and incorporating control mechanisms across all stages of development, deployment, and operation will become a fundamental requirement for future enterprise AI adoption.