Questions this engagement should answer
- Which decisions remain human-authorized and which actions may be automated?
- Are identity, secrets and permissions bounded outside the model?
- Can every action be attributed, reviewed, stopped and recovered?
- How are tool output, retrieved content and user messages treated as untrusted input?
What the work includes
- Architecture and trust-boundary review for the agreed workflow
- Inventory of tools, identities, permissions, secret paths and approval points
- Abuse-case review covering prompt injection, excessive agency and data exposure
- Recommendations for staged access, observability, stop conditions and rollback
Deliverables
- AI-agent access and authorization map
- Threat scenarios and control-gap register
- Proposed permission tiers and approval boundaries
- Pilot acceptance criteria, monitoring requirements and shutdown plan