If an AI agent is attesting your controls, who’s attesting the agent?

In the world of cybersecurity, AI agents are increasingly being used to attest to the effectiveness of controls. However, this reliance on automated attestations raises the question of who is attesting to the AI agents themselves. Traditional compliance-first approaches are struggling to keep pace with the rapid evolution of AI systems.

Many security teams are caught in a cycle of proving they have controls in place rather than actively reducing actual risks. This gap between appearing secure and being secure is further exacerbated by the use of AI agents, which can automate the appearance of control just as easily as they can automate the control itself. The challenge lies in maintaining trust when AI starts making or influencing security decisions.

To address this, organizations must focus on visibility, establishing a complete inventory of all AI systems, model APIs, and embedded AI components. Every deployed AI agent should be treated as a new actor in the environment, with its own identity, scoped access, and proper authorization. High-impact actions should always retain meaningful human oversight.

Ongoing enforcement through runtime controls such as least-privilege access, just-in-time permissions, and automated containment is crucial to minimizing the impact of any potential security incidents. While automation can alleviate some of the burden, it should not be the primary focus. Instead, the goal should be a different target: continuous trust through aligned risk signals, controls, and evidence as the environment evolves.

The key lies in creating governance programs that make the approved path easier than the shadow one while maintaining meaningful human oversight. By adopting this approach, organizations can better navigate the evolving landscape of AI-driven security and ensure their controls are not just theoretical, but truly effective.