attacks/taxonomy.json

{
  "version": "1.0",
  "attack_types": {
    "A1": {
      "name": "Direct Injection",
      "description": "Plaintext malicious instructions embedded directly",
      "surface": "tool.description",
      "example_intents": [
        "instruction override",
        "data exfiltration",
        "privilege escalation"
      ]
    },
    "A2": {
      "name": "Unicode Steganography",
      "description": "Zero-width characters, homoglyphs, and invisible Unicode to hide payloads",
      "surface": "tool.description",
      "example_intents": [
        "hidden commands",
        "invisible instructions",
        "obfuscated payloads"
      ]
    },
    "A3": {
      "name": "Base64 Payload",
      "description": "Base64-encoded instructions designed to be decoded and executed",
      "surface": "tool.description",
      "example_intents": [
        "encoded exfiltration",
        "obfuscated commands",
        "hidden scripts"
      ]
    },
    "A4": {
      "name": "Schema Poisoning",
      "description": "Malicious parameter names, types, or descriptions in tool input schemas",
      "surface": "tool.inputSchema",
      "example_intents": [
        "credential theft",
        "path traversal",
        "parameter injection"
      ]
    },
    "A5": {
      "name": "Output Injection",
      "description": "Injection attacks embedded in tool return values",
      "surface": "tool.output",
      "example_intents": [
        "fake system alerts",
        "credential phishing",
        "destructive commands"
      ]
    },
    "A6": {
      "name": "Cross-Server Shadowing",
      "description": "Attempts to override or shadow legitimate tools from other servers",
      "surface": "tool.description",
      "example_intents": [
        "tool impersonation",
        "priority hijacking",
        "sandbox bypass"
      ]
    },
    "A7": {
      "name": "Rug Pull",
      "description": "Tool definitions that change behavior after initial approval",
      "surface": "tool.description",
      "example_intents": [
        "post-approval telemetry",
        "fake updates",
        "feature injection"
      ]
    },
    "A8": {
      "name": "Error Message Injection",
      "description": "Malicious instructions embedded in error messages",
      "surface": "tool.error",
      "example_intents": [
        "error-driven tool calls",
        "credential phishing",
        "supply chain attacks"
      ]
    },
    "A9": {
      "name": "Nested Encoding",
      "description": "Multi-layer encoding chains to evade detection",
      "surface": "any",
      "example_intents": [
        "double encoding",
        "mixed encoding",
        "encoding chains"
      ]
    },
    "A10": {
      "name": "Semantic Camouflage",
      "description": "Malicious instructions disguised as legitimate documentation",
      "surface": "tool.description",
      "example_intents": [
        "fake best practices",
        "compliance pretexting",
        "documentation-wrapped attacks"
      ]
    },
    "A11": {
      "name": "Sampling Exploitation",
      "description": "MCP sampling requests hijacking LLM completions",
      "surface": "tool.output",
      "example_intents": [
        "resource theft",
        "conversation hijacking",
        "covert invocation"
      ]
    },
    "A12": {
      "name": "Preference Manipulation",
      "description": "Subtly alters tool ranking so AI prioritizes malicious tools",
      "surface": "tool.description",
      "example_intents": [
        "priority hijacking",
        "traffic interception",
        "default replacement"
      ]
    },
    "A13": {
      "name": "Parasitic Toolchain",
      "description": "Chained tools escalating attacks through interlinked network",
      "surface": "tool.description",
      "example_intents": [
        "pipeline injection",
        "middleware interception",
        "multi-stage exfil"
      ]
    },
    "A14": {
      "name": "Supply Chain .pth Poisoning",
      "description": "Python .pth files for persistent code on interpreter startup",
      "surface": "tool.description",
      "example_intents": [
        "persistent backdoor",
        "startup injection",
        "dependency poisoning"
      ]
    },
    "A15": {
      "name": "Indirect Content Injection",
      "description": "Poison in fetched content not in tool definition",
      "surface": "tool.output",
      "example_intents": [
        "content injection",
        "document poisoning",
        "hidden instructions"
      ]
    },
    "A16": {
      "name": "System Prompt Leakage",
      "description": "Tool definitions crafted to extract the host LLM's system prompt",
      "surface": "tool.description",
      "example_intents": [
        "prompt extraction",
        "system instruction leak",
        "alignment verification pretext"
      ]
    }
  }
}