When AI Agents Can Act, Security Must Govern Their Actions

AI agents bring a practical security question into everyday operations: what should an agent be allowed to do on someone’s behalf?

Investigating an issue may require reading logs and inspecting a system. Resolving it may require changing configuration, restarting a service or modifying access. Each step carries a different level of risk.

Useful autonomy depends on clear boundaries

Our view is that useful autonomy depends on clear boundaries:

• Give each agent access appropriate to its task.
• Define which actions can proceed automatically.
• Require approval when an action carries greater consequences.
• Preserve evidence of the request, authorization, action and result.
• Verify whether the intended outcome was achieved.

Security for agents and the work they perform

This applies both to agents performing security work and to the protection of agents themselves. Security teams need to understand what agents can reach and retain control over what they can change.

As organizations expand their use of agents, these controls should become part of how work gets done.