Cognitive

Part 7 · Agentic AI Security

ASI10: Rogue Agents

An agent can drift from its purpose while each individual action still appears legitimate.

3 min readBuilds on ASI01: Agent Goal Hijack, ASI08: Cascading Failures

OWASP ASI10. This material is for defense. Only test systems you own or are authorized to test.

A rogue agent can spread through valid access while each step looks normal.

What it is

Rogue agents are malicious or compromised AI agents that deviate from their intended function or authorized scope, acting harmfully, deceptively, or parasitically within multi-agent or human-agent systems.

Their individual actions may look normal, which creates a containment gap for traditional rule-based controls.

How the failure unfolds

Behavior monitoring looks at patterns over time, and containment limits what a drifting agent can reach.

Figure 49.1. Behavior monitoring looks at patterns over time, and containment limits what a drifting agent can reach.

Common patterns

  • Goal drift: behavior moves away from the assigned objective.
  • Scope expansion: the agent acquires access it was not given for the task.
  • Self-replication: it creates persistent copies or tasks.
  • Concealment: logs or reports hide what it is doing.
  • Reward hacking: it games the metric used to judge success.

Defenses

Monitoring and containment limit a drifting agent.
  1. Keep signed audit logs. Record actions, tool calls, and communication.
  2. Create trust zones. Restrict communication between environments.
  3. Monitor behavior. Compare activity with an approved baseline.
  4. Contain quickly. Provide kill switches and credential revocation.
  5. Attest identity. Verify each agent and its configuration.
  6. Require re-approval. Review agents before restoring access.

Common misconception

Common belief

"A rogue agent must have been hacked."

In practice

Misalignment and poor objectives can cause an agent to go rogue with no outside attacker.

Key takeaways

  • Rogue behavior can emerge with or without an outside attacker.
  • Monitor patterns of behavior over time.
  • Containment requires isolation and a way to revoke access quickly.