A rogue agent may look normal while it quietly exfiltrates data, keeps unauthorized access, or works against its operator's goals. Because agents are autonomous and long-running, detection and fast containment matter as much as prevention.
Continuously monitor agent behavior against an expected baseline, enforce policy outside the model, keep immutable audit logs, and provide kill switches that revoke an agent's credentials and stop its tasks immediately.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2026Latest
#10asi10-2026 Rogue Agents
