Excessive agency occurs when LLM systems are granted autonomy beyond their reliability and control maturity, allowing cascading failures across connected systems. Attackers exploit autonomy boundaries to amplify impact from a single prompt or model misclassification. The challenge is governance of action authority, not just model output quality.
Define strict autonomy boundaries with risk-tiered controls and mandatory human oversight for critical actions. Limit multi-step autonomous execution and enforce checkpoints at sensitive transitions. Instrument full action traceability and rollback paths to contain failures quickly.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2024Latest
#08llm08-2024 Excessive Agency
