Securing the Future of Autonomous Systems Through Advanced Identity Governance

The rise of agentic artificial intelligence offers businesses a chance to fundamentally reinvent their operations and boost efficiency. However, these autonomous systems present significant hurdles for technical leadership. Organizations face a difficult balance: they must rapidly adopt these tools to remain competitive in revenue and productivity, yet they must also manage a technology that changes weekly and presents massive, often unquantifiable security risks. According to Matt Caulfield, Cisco’s VP of Product for Identity, many companies currently have AI agents functioning without clear ownership or proper access limitations. Developers frequently link these agents to core production environments without involving IT departments, creating dangerous security gaps.

Autonomous agents differ from traditional software because they can independently query data, initiate complex workflows, and make rapid decisions without human oversight. This independence renders current identity and access management (IAM) frameworks insufficient. While human identities are typically established during hiring and updated infrequently, agents often operate using broad, inherited credentials from their environment. Previous tools designed for non-human entities focused on static API keys and service accounts, failing to provide the granular, action-by-action oversight that dynamic AI agents require. Consequently, business adoption of these agents is currently outpacing the development of governance strategies.

To address this, Cisco is applying Zero Trust principles to the world of autonomous agents. This strategy is built on three main pillars: total visibility, clear accountability, and precise enforcement at the level of individual actions. Visibility is the primary challenge, as "shadow agents" often operate undetected after being deployed for temporary developer tasks. By utilizing its existing network infrastructure, Cisco’s Duo Agentic Identity can identify these hidden agents the moment they begin communicating across a network, providing a live inventory rather than a static scan.

Once an agent is identified, it must be governed through a full lifecycle, similar to a human employee. This involves assigning every agent a specific human sponsor and treating the agent as a distinct identity within the directory. By supporting protocols like the Model Context Protocol (MCP) and OAuth 2.1, the system remains neutral to specific AI vendors. Security is further tightened by restricting agents to only the necessary tools for a specific task. Through an MCP gateway, the system can block high-risk actions—such as exporting a database rather than just reading it—in real-time. Establishing these governance frameworks now is essential for businesses to safely achieve the large-scale benefits of AI automation.

---

Read original at ZDNet AI.