Skip to main content

Red Team

Adversarial red-teaming for AI governance. A curated, downloadable library of named attack patterns — prompt injection, encoding-evasion, RAG contamination, destructive tool calls — plus the tooling to fire each one against your Guardians and verify the defense in your own tenant.

What it is​

Red Team is the offensive half of the safety loop. The Threat Library catalogs named adversarial patterns with the expected Guardian verdict and a downloadable JSON fixture you can drop into a Test Suite. The Teleological Data Generator (TDG) auto-generates syntactic, semantic, and malicious variations of each pattern so a single named attack expands into a full adversarial suite.

Why it matters​

  • Test the defense, not the dashboard. Each threat ships a fixture you run against your Guardian in your tenant — you verify the block, not a vendor claim.
  • Auto-expanding attack surface. TDG turns one named pattern into n adversarial variations (encoding swaps, intent attacks, schema violations), so your suite grows with the threat landscape without manual authoring.
  • Closed loop. A red-team run that finds a gap feeds straight back into Guardian training and policy ingestion.
  • Reproducible evidence. Each red-team run mints a signed eh_ receipt. "This Guardian blocks Threat T-014 and all 47 TDG variations" is a statement a regulator can re-run, not a dashboard claim.

How it works​

  1. Pick a threat from the Threat Library — each entry includes the attack, the expected verdict, and a fixture.
  2. Expand it with TDG variations (or use it as-is).
  3. Run the fixture against your Guardian via the Simulation or Evals surface.
  4. Verify the Guardian returns the expected corrected / blocked verdict.
  5. Feed gaps back as new training scenarios.

Get started​

→ Threat Library — named patterns + downloadable fixtures. → Testing & Simulation — TDG and scenario generation. → Evals: turn red-team runs into a signed, replayable receipt.

Last updated on