Due Diligence Report: Agentic AI Deployment Liability
An assessment of what an autonomous agent is actually authorized to do, versus what its designers assumed.
Executive Summary
This assessment evaluated an agentic AI deployment adversarially, focused specifically on undefined authority boundaries rather than code correctness, which the deployment had already passed through standard review.
Findings
The deployed agent held write access to a production data system with no formally documented limit on what records it was authorized to modify.
Recommendation: Implement explicit, enforced permission boundaries before further scaling the deployment; do not rely on prompt-level instruction alone.
The agent had been validated against expected inputs but not against adversarial or malformed inputs designed to probe boundary conditions.
Recommendation: Commission a dedicated adversarial test pass before production scaling, structured like a penetration test rather than a functional QA pass.
No formally tested procedure existed to immediately revoke the agent's access in the event of anomalous behavior.
Recommendation: Build and test a rapid revocation procedure as a standing operational requirement, not an incident-time improvisation.
Conclusion
The liability in this engagement was not in the model's outputs, it was in the authority boundary around those outputs. This is a governance and systems design finding, not a model-quality finding, and it required a different team and methodology to surface than the deployment's original code review.
Frequently Asked
What is the biggest blind spot in agentic AI governance?
Undefined write-access and authority boundaries. Standard code review validates correctness, not what an agent is actually permitted to do once deployed.