AI Red-Team Testing

You should be able to attack your own governance before an adversary does. EVE ships a customer-runnable adversarial regression suite — single-step cases and multi-step chains — that you run against your own configuration. Blocked attacks provably cause zero unauthorized side effects, and the results are signed.

What the harness does

  • Adversarial suites — single-step attack cases plus multi-step chains that try to slip a prohibited outcome past governance.
  • Zero unauthorized side effects on blocked attacks — the reference run recorded 16 cases plus 2 chains with 0 unauthorized side effects.
  • Signed reports — results are captured in a signed red-team report.
  • Safe by default — local recording targets, no destructive tests, dry-run execution.

Evidence example

signed red-team report { passed, unauthorized_side_effects: 0 }

SDK context

The red-team suite runs through the same governance path your agents use, so a blocked attack never reaches the underlying tool.

Readiness

The red-team harness is PILOT_READY and customer-pilot ready at supported boundaries (Python; embedded-service and sidecar modes), covered by the artifact-scan phase report (16 cases + 2 chains, 0 unauthorized side effects).

Limitations

  • The suite is representative, not exhaustive — customers add their own fixtures for their own risks.
  • Standards mappings that accompany the harness are descriptive and are not a certification.
  • It runs safe by default (local recording targets, no destructive tests, dry-run); wiring it against a live target is a deliberate configuration step.

Next step

Run the packaged adversarial suite against your own configuration in a pilot and keep the signed report.

Part of the EVE AI Core control plane Deterministic AI Governance Control Plane → Policy decisions that return the same result for the same input every time, before execution.