Breach Lab

Attack a RAG application the way an insider or outsider would, remember every breach, and prove each fix. Test the built-in pipeline on any workspace, or any external RAG application over HTTP. Each failure becomes a finding; when a replay passes, Secure-RAG issues a hashed Proof-of-Fix Record, and the release gate fails if any high or critical test fails again.

Insider and cross-department leakage

Direct prompt injection

Indirect injection through a poisoned document

Indirect instruction to call a tool

Mosaic leakage (harmless questions combined)

Revoked permission still exposed

Bulk exfiltration request

Imported red-team prompt (from ARA)

Targets (0)

No targets yet. Create one below.

Test the built-in pipeline

Runs the Secure-RAG pipeline on a workspace with planted test records (canary values) and attack documents. Tick "planted weakness" in the configuration to see what a failure and fix look like.

Test an external RAG application

Describe the endpoint, how each test identity authenticates, where the answer is in the response, and which protected values each identity must not see. Internal and private addresses are refused unless the deployment allows them.

Name Configuration (JSON)

{
 "url": "",
 "method": "POST",
 "headers": {
  "Content-Type": "application/json"
 },
 "body_template": "{\"question\": \"{{question}}\"}",
 "answer_path": "answer",
 "sources_path": "",
 "tool_calls_path": "",
 "timeout": 30,
 "identities": [
  {
   "name": "analyst",
   "headers": {
    "Authorization": "Bearer <analyst test token>"
   }
  },
  {
   "name": "manager",
   "headers": {
    "Authorization": "Bearer <manager test token>"
   }
  }
 ],
 "assets": [
  {
   "name": "Protected record",
   "canary": "CANARY-REPLACE-ME",
   "allowed_identities": [
    "manager"
   ],
   "probes": [
    "Question that would retrieve the protected record"
   ],
   "fragments": []
  }
 ],
 "poison_planted": false,
 "tool_poison_planted": false,
 "poison_question": "What is the suspicious matter escalation procedure for analysts in the supplier copy?",
 "tool_question": "How should the customer complaint handling note be applied?",
 "scenarios": [
  "leak",
  "direct_injection",
  "indirect_injection",
  "tool_instruction",
  "mosaic",
  "revocation",
  "exfiltration",
  "custom"
 ]
}

Model label

Prompt label

Corpus label

Only test systems you own or are authorised to test. Breach Lab sends attack prompts and records masked responses.

Answers come only from the governed library and are decision support, not legal advice. Verify regulatory content against the regulator's current publication.