Agentic AI is emerging as a primary workload within an enterprise’s AI environment. The question now is whether those agents will do what you intend for them to do and not go rogue to create new security threats. Agentic AI compounds the identity control and governance challenge that organizations face, and with agents creating agents, organizations should consider capabilities that can help to manage and support business needs effectively while ensuring security is robust.
Agentic AI may not be a primary AI workload today, but it will be soon (with inference retaining pole position, of course). In a perfect world, these agents will conduct tasks exactly as planned. But it isn’t a perfect world, and there have already been instances where agents have gone rogue, developed their own sense of mission, and turned into attackers. But even when the agent does not possess malicious intent, control and visibility remain paramount, not just for security but also for optimizing token economics, monitoring the span of control, and maintaining business continuity.
Organizations are now tasked with developing and deploying secure agentic AI solutions with robust built-in security features that can be governed properly.
Agents are not users
AI may be one of the greatest tools available for cybersecurity. At the same time, AI can be one of the most dangerous tools used as a weapon against critical IT resources. “The uncomfortable truth is that if you can use AI to find vulnerabilities in your own products, the bad guys can too,” Jon Green, CTO, Networking, HPE, wrote.
So, how do you build agents with robust built-in security features, and which can be governed well? How do you implement guardrails that excel at monitoring the behavior of agents? For agentic AI, the fundamental premise is that you need to have control over key elements such as code execution, data access, and API calls—and that requires a well-developed security foundation.
Agents have identities, but they aren’t the same as users. The behavior of agents can’t be governed with the same type of processes and controls as humans. Agents act on behalf of the user. And yes, sometimes these agents go rogue and escalate privileges and tasks in ways developers never anticipated, but that is only one type of security risk that needs to be addressed.
When developed right, each agent runs with a set of credentialed controls. That set of identity controls should limit the potential of sprawl or additional rogue behaviors—if the protections are implemented flawlessly. But that is very difficult to do.
Challenges in developing secure agents
Many digital estates have become so complex and sprawling that managing AI security effectively is difficult — even in the best-case scenarios.
Security protocols start with agent development. There are a number of agent-building technologies such as that programmers use to build the code that will execute as an agent. Then, you need a sandbox or clearing house to send the agent to review and fix any flaws in the code. The development process will also include a policy function that directs the assignments and defines the guardrails of the agent, which, in theory, should prevent it from rogue tasks.
However, even when thoughtfully programming an agent with legitimate intent can still operate in nefarious ways. That’s why agents’ behavior needs to be monitored by multiple layers of security tools throughout the IT infrastructure — what security professionals call “defense in depth”. IT and security teams need to be able to have visibility across the entire IT estate. This includes monitoring and detection at the network level because that’s where near real-time agent activity is visible. The network has a front row seat to monitor agent behavior and alert on rogue activity.
Another governance risk is shadow AI. Here, it is less the agents going rogue than users accessing easy-to-use agent-building functions outside IT purview and processes. Unsanctioned development and use allow AI agents to move through the system without oversight, making it more difficult for the agent security solutions to protect the organization.
Third-party agents
Autonomous agents and shadow IT are not the only risks. Third-party agents are also a security threat because they, by definition, come from outside the organization and its security controls. Even if it was built specifically for the company, a third-party agent may or may not conform to the organization’s security standards or is running on infrastructure the organization doesn’t own.
While working to secure third-party agents for your organization, the process should be held to the same rigorous security testing and validations that internal agents go through so that they have the same guardrails and governance as internal agents.
Conclusion
Guardrails set up the boundaries and constraints for AI agents, but comprehensive security requires a defense-in-depth that offers visibility and monitors the agent on every step of its path. But visibility is only the first step. The IT infrastructure should also enforce policies, limit privileges, and help protect data when standard agent security controls are not enough. The security ecosystem, including IT infrastructure, including servers, storage, and networking, should work together to help provide a more comprehensive approach to agentic security
