WeaknessASC-2026-0010Major weakness
Rogue agents
HighOpen, no known fix
Agents drift from intended behavior, through compromise or misalignment, and act outside policy without detection.
How to detect it
- No action-level audit logging
- No behavior baselines or anomaly alerts
- The agent can modify its own configuration or tools
Workarounds
- Log every tool call with inputs, outputs and identity
- Alert on out-of-policy actions and drift from a behavior baseline
- Make agent configuration immutable at runtime and keep a kill switch
There is no reliable way to tell a compromised or misaligned agent from its outputs alone. Contain and monitor; revisit as detection research matures.
How to investigate
- Build a timeline of the agent's actions and compare it with its baseline
- Check for configuration, prompt or tool list changes before the drift
- Quarantine the agent, revoke its credentials and review everything it touched