Domain 5 · AI, RAG, Model, and Agentic Security
AI Agent Trust and Prompt-Injection Analysis
Model the agent trust graph — tool, data and model permissions, autonomy, persistence, approvals, external communication and code execution — and the path from untrusted input through the model and a tool to a privileged action.
Decision it answers
Which agents exist, what can they reach, and what is the maximum blast radius if one is manipulated?
Inputs it needs
- Agent inventory with assumed identities
- Tool and connector permission scopes
- Retrieval sources, corpora and vector stores
- Approval and autonomy configuration
Methods applied
- Agent trust-graph construction
- Untrusted-input to privileged-action tracing, focused on trust boundaries rather than prompt filtering alone
- Irreversibility weighting of available actions
- Autonomous action budget and approval-threshold modelling
Outputs produced
- Agent blast-radius statement
- Trust-boundary violations
- Actions requiring human approval on irreversibility grounds
- Residual uncertainty and validation status
Must remain human-owned
Authorization, acceptance of residual risk, containment decisions, disclosure and notification, and any communication to a board or regulator. No analytical capability creates operational authority.