Skip to content

Free for nonprofits, NGOs, think tanks, and institutes. Grant funded by James Scott, administered by the Embassy Row Project.

ArtOfTheHack home

9 minute read

Governing Autonomous Agents in Security Operations

Agents are entering security operations faster than the controls that should bound them. Authority, reversibility, and evidence are the prerequisites.

The gap between capability and control

Agents can now triage, enrich, correlate, and in many products act. What has not kept pace is the control surface: what an agent is permitted to do, on whose authority, with what evidence, and what happens when it is wrong.

Four conditions for agent action

An agent action in a production security environment should satisfy every one of these conditions before it executes. If any is unmet, the correct behavior is to prepare the action and escalate rather than proceed.

  • Bounded authority: an explicit scope of systems, actions, and thresholds
  • Evidence sufficiency: the supporting evidence meets a stated standard
  • Reversibility: a rollback path exists, or a human approves the irreversible step
  • Recorded basis: the full decision object is written before the action executes

Agents as a protected surface

Agents are also a target. Prompt injection, tool abuse, credential inheritance, and memory poisoning are live attack paths against the automation itself. Agent governance and agent security are the same program viewed from two directions.