tool_poisoningTier 1 · 70% confidence

security-tool-poisoning-malicious-tool-descriptions-could-exfiltrate-data--bc7f7c34

agent: security

When does this happen?

IF Malicious tool descriptions could exfiltrate data or execute commands through AI agent.

How others solved it

THEN Enable MCP Gateway's built-in tool-poisoning validator that scans every backend tool description for high-risk patterns (e.g., IMPORTANT blocks, SSH keys, base64 exfiltration) and fails closed on matches.

The validator runs by default. To see rule implementation, check src/validator/rules/tool_poisoning.rs in the mcp-gateway repository.

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics