Assessment line 03 · Models, autonomous agents and production risk
AI Governance and Security for systems that have moved beyond a chat interface.
When AI models take actions, read production databases or write to third-party APIs, standard software security checks miss the risk. We audit the data lineage, the model evaluation set, the agent permissions and the human-in-the-loop controls.
3 to 5
Weeks from scoping call to worklist
6
Areas covering models, agents, practices and governance
Every
Agent mapped for what it can reach and what it can trigger
Ours
Evaluation set built by us, never supplied by the vendor
The exposure map
An AI system is a data path with a decision in the middle and an agent on the end.
We map the whole path rather than testing the model on its own, because in practice the damage happens at the joins and at the point where something acts.
Figure 1 · Where exposure sits across an AI systemAuditTrax Labs
Exposure point found in a typical first assessment Governance layer
What we examine
Six areas covering the model, the agent, the data and the team.
01
Inventory of everything running
Every model, copilot, agent and API call in production, found through code, cloud accounts and expense records rather than by asking people to self declare.
02
Agent permissions and blast radius
For every agent that acts: what it can read, what it can write, what it can trigger, and what the worst outcome is if someone talks it into doing the wrong thing.
03
Model provenance and licence
Where the model came from, what it was trained or tuned on, and whether you hold the rights to use it the way you are using it.
04
Data paths and leakage
Every route a customer record can take once it enters a prompt: into the index, into logs, into a third party processor, into a jurisdiction you did not intend to be in.
05
Adversarial testing
Prompt injection through content the system reads, jailbreaks against the guardrails, tool permission abuse, and poisoned or unpinned dependencies.
06
Practices and governance
How your teams actually build and change these systems: approval before deployment, change control on prompts and thresholds, and incident logging.
Tiers and pricing
Three tiers, fixed fee, agreed before work starts.
Tier 1 · Inventory & Board Readiness
Screening
Find everything running and establish the board-level exposure.
$26,000
- Shadow AI discovery across code & expenses
- System and agent register
- High level blast radius per agent
- Priority ranked worklist
Tier 2 · Most chosen
Full Assessment
Complete evaluation across all four attack surfaces.
$52,000
- Everything in Screening
- Detailed permission & blast radius mapping
- Model provenance & licence review
- Data leakage path tracing
Tier 3 · High Exposure
Full + Adversarial
Hands-on injection, jailbreak and permission testing.
$78,000
- Everything in Full Assessment
- Prompt injection & jailbreak testing
- Held-out evaluation benchmark run
Questions we get asked
The awkward ones, answered first.
Our vendor will not disclose what the model was trained on.
That refusal is itself a finding and we write it as one. We then assess what can be established from outside, and state plainly what the gap means for the risk you are carrying and who is holding it.
Is this a penetration test?
Tier 3 includes adversarial testing, but the engagement is wider. A penetration test asks whether someone can get in. This asks whether the system should be trusted with the decision you have handed it.
We built it ourselves, so is it simpler?
Usually harder. Access is better and documentation is worse, and the evaluation set is often missing entirely because the team was shipping under pressure.
Will you fix what you find?
Not on the same engagement. Any connected firm that could do remediation is disclosed on page one, and the assessor takes no part in that work.
AI Governance and Security
Find out what is actually running before someone else does.
Discovery is the first week of an AI governance assessment, and it is usually the week that changes the conversation.