Agent Assurance Audit
Audit the controls around an AI agent and test its behaviour across the full loop — understand, observe, locate, act, verify.
For: Organisations deploying or buying agents that take actions on their behalf.
- Ownership, decision rights and stop authority
- Irreversible actions and human confirmation
- Tool permissions and least privilege
- Step-level logging, replay and audit trail
- Held-out task completion across the agent loop
- Direct and indirect prompt injection
- Tool-failure, recovery and hand-off behaviour
- Findings register with severity and owners
- Scorecard: Safe · Skilled · ROI-Fit
- Agent-loop trace analysis
- Release-gate verdict: Go · Conditional · No-Go
- Prioritised recommendations
We audit the loop, not just the answer
Every agent understands, observes, locates, acts and verifies. Each stage fails in its own way, so each stage is tested on its own.
Understand
Interprets the request, the goal and the constraints.
We test: Ambiguous, adversarial and out-of-scope requests.Observe
Gathers context: conversation, permissions, system state.
We test: Missing, stale or manipulated context — including hidden instructions.Locate
Finds the right data, document or tool for the next step.
We test: Wrong-record retrieval, permission leaks, tool selection.Act
Calls the tool, writes the record, sends the message.
We test: Irreversible actions, argument correctness, least privilege.Verify
Checks the result; recovers or hands over when it fails.
We test: Tool failures, silent errors, loops and escalation behaviour.Scoped to the work, fixed before it starts
Price and length depend on the volume of the work: the number and complexity of the systems in scope, the access we need and the depth of testing. Both are evaluated during scoping and fixed in writing before anything begins. How engagements work →