{
  "id": "invariant-mcp-tool-poisoning-2025-04",
  "name": "MCP Tool Poisoning Attacks (Invariant Labs)",
  "date_reported": "2025-04-01",
  "researcher": "Luca Beurer-Kellner, Marc Fischer",
  "org": "Invariant Labs",
  "sources": [
    {
      "url": "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
    }
  ],
  "magicsword_coverage": [],
  "borderline": false,
  "borderline_note": null,
  "ai_role": "ai_agent_or_mcp_as_target",
  "ai_role_secondary": null,
  "ai_role_detail": null,
  "technique_summary": "Malicious MCP tool descriptions hide instructions only the model sees (e.g., read ~/.ssh/id_rsa and mcp.json into sidenote params). Also demonstrated rug pulls (change description after approval) and cross-server shadowing that redirects trusted send_email tools to attacker addresses. Affects Cursor and other MCP clients.",
  "initial_access": "User connects agent to malicious or rug-pulled MCP server",
  "ai_tooling_and_models": [
    "MCP tool descriptions",
    "Cursor and other MCP clients"
  ],
  "other_tooling": [],
  "victims_and_sectors": "unknown",
  "regions": [
    "unknown"
  ],
  "attribution_or_cluster": "unknown",
  "scale_and_metrics": "unknown",
  "prevention_angle": "Show full tool descriptions; pin tool hashes; isolate MCP servers; least privilege",
  "related_magicsword_projects": [
    "LOLSkills"
  ],
  "source_notes": null,
  "category": "attacks-on-ai",
  "ai_involvement_confidence": "confirmed",
  "last_reviewed": "2026-10-20",
  "attack_flow": {
    "nodes": [
      {
        "id": "step-1",
        "title": "A malicious MCP server is connected",
        "description": "The user adds a server or trusts one that later changes.",
        "kind": "access",
        "source_urls": [
          "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
        ]
      },
      {
        "id": "step-2",
        "title": "Tool descriptions carry hidden instructions",
        "description": "Instructions are visible to the model and may shadow another server’s tools.",
        "kind": "ai",
        "source_urls": [
          "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
        ]
      },
      {
        "id": "step-3",
        "title": "Agent reads files or redirects a tool",
        "description": "The demonstrations induce sensitive-file reads and manipulate email destinations.",
        "kind": "execution",
        "source_urls": [
          "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
        ]
      },
      {
        "id": "step-4",
        "title": "Sensitive data is sent to the attacker",
        "description": "Tool arguments or redirected operations carry information across the trust boundary.",
        "kind": "impact",
        "source_urls": [
          "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
        ]
      }
    ],
    "edges": [
      {
        "source": "step-1",
        "target": "step-2"
      },
      {
        "source": "step-2",
        "target": "step-3"
      },
      {
        "source": "step-3",
        "target": "step-4"
      }
    ]
  },
  "atlas_techniques": [
    {
      "id": "AML.T0051.001",
      "rationale": "Tool descriptions contain indirect model instructions.",
      "source_urls": [
        "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
      ]
    },
    {
      "id": "AML.T0109",
      "rationale": "A trusted server can change its behavior after approval.",
      "source_urls": [
        "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
      ]
    },
    {
      "id": "AML.T0086",
      "rationale": "The demonstrations exfiltrate through tool arguments or redirected tools.",
      "source_urls": [
        "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
      ]
    }
  ]
}