How TrustAI works
Reconstruct every agent.
Point TrustAI at the hub; it captures how each agent behaves.
TrustAI walks the MCP hub and enumerates every agent on it, with the tools, prompts, and authorizations each one was handed. No code changes to the agents.
TrustAI Assessment
Enumerate every Joule agent on the S4H PRD hub and prepare them for the control battery
MCP hub connected · 5 agents enumerated, tools and prompts captured
Agents reconstructed for assessment
| Agent | Business Scenario | Model | Status |
|---|---|---|---|
| Core HR Assistant | HCM · Employee Central | claude-haiku-4-5 | Reconstructed |
| Financial Closing Assistant | FIN · Financial Closing | claude-haiku-4-5 | Reconstructed |
5 agents on the hub · next: 13 preregistered controls against each
Run the preregistered battery.
Every control area your auditors ask about, tested against the reconstructed agent.
Where the agent's data comes from, where its outputs go, and whether retention, deletion, and consent obligations hold. Regulated data is flagged before any of it lands in a model's context.
Control battery
Dispute Resolution Agent · 13 preregistered controls
10/13passed
Every rate carries a Wilson 95% confidence interval · acceptance bars preregistered before the run
A verdict you can defend.
Backed by the evidence behind every finding.
A clear verdict, whether that's safe to deploy, deploy after remediation, or don't deploy, with the statistical evidence behind it.
Verdict
Dispute Resolution Agent
Do not deploy as scoped
FAIL
CI gate @ 5.0%
Pooled attack-success 14.2% (95% CI 11.3–17.6%, n=480), so the 5.0% gate on the confidence-interval upper bound does not clear. Remediate grounding and adversarial failures, then re-assess.
