config.fixed-attacks.example.json

{
  "target": {
    "baseUrl": "https://YOUR-ENDPOINT-HERE.example.com",
    "agentEndpoint": "/v1/chat/completions",
    "applicationDetails": "Replace baseUrl/agentEndpoint with your target. This config replays a set of FIXED attacks (proven breaches in data/fixed-attacks/nvidia-ai-safety.json) verbatim against the target, AND — because customAttacksOnly is left off — also runs the normal category/strategy LLM-generated attacks alongside them. All attacks are judged the same way and appear in one report.",
    "policyFile": "policies/safety-classifier.json",
    "customApiTemplate": {
      "headers": { "Content-Type": "application/json" },
      "bodyTemplate": "{\"model\": \"YOUR-MODEL\", \"messages\": [{\"role\": \"user\", \"content\": \"{{message}}\"}], \"max_tokens\": 512, \"temperature\": 0}",
      "responsePath": "choices[0].message.content"
    }
  },
  "requestSchema": {
    "messageField": "message",
    "roleField": "role"
  },
  "responseSchema": {
    "responsePath": "choices[0].message.content",
    "toolCallsPath": "choices[0].message.tool_calls",
    "userInfoPath": "",
    "guardrailsPath": ""
  },
  "auth": {
    "methods": ["none"],
    "credentials": [],
    "apiKeys": {}
  },
  "customAttacksFile": "data/fixed-attacks/nvidia-ai-safety.json",
  "attackConfig": {
    "adaptiveRounds": 1,
    "concurrency": 1,
    "delayBetweenRequestsMs": 100,
    "enableLlmGeneration": true,
    "includeSeedAttacks": false,
    "enableMultiTurnGeneration": false,
    "enableAdaptiveMultiTurn": false,
    "maxAdaptiveTurns": 1,
    "maxMultiTurnSteps": 1,
    "strategiesPerRound": 10,
    "attacksPerStrategy": 2,
    "categoryParallelism": 1,
    "multiTurnGenerationRate": 0,
    "skipIrrelevantCategories": false,
    "requireReviewConfirmation": false,
    "llmProvider": "openrouter",
    "llmModel": "anthropic/claude-sonnet-4.5",
    "judgeProvider": "openrouter",
    "judgeModel": "anthropic/claude-sonnet-4.5",
    "enabledCategories": [
      "indirect_prompt_injection",
      "prompt_injection",
      "output_evasion",
      "content_filter_bypass",
      "encoding_serialization_attack"
    ],
    "enabledStrategies": [
      "critical_deadline_pressure",
      "educational_series_framing",
      "academic_computation_disguise",
      "role_consistency_override",
      "trusted_source_context"
    ]
  }
}