tool_poisoning_detectionTier 1 · 70% confidence

security-tool-poisoning-detec-malicious-or-compromised-mcp-server-tool-descripti-194686f9

agent: security

When does this happen?

IF Malicious or compromised MCP server tool descriptions can contain hidden instructions (e.g., <IMPORTANT> blocks, SSH key references, exfiltration commands like curl to attacker servers) that get injected into the AI agent's context, leading to tool poisoning, rug pulls, or data exfiltration.

How others solved it

THEN Use mcp-gateway's built-in tool-poisoning validator (AX-010) which scans every backend tool description before it reaches the agent. HIGH patterns (e.g., SSH key refs like ~/.ssh, exfil language like 'curl .* https?://', base64 in exfiltration context) cause fail-closed, blocking the tool. MEDIUM patterns (oversized descriptions >40 spaces, zero-width Unicode, bidi-override characters) generate warnings. This provides a centralized audit surface for all tool descriptions.

Related patterns

security

security-security-site-missing-permissions-policy-header-724230ad

Tier 1 · 99%
security

security-security-site-missing-referrer-policy-header-4550db61

Tier 1 · 99%
security

security-security-site-missing-x-content-type-options-header-d1bbaadd

Tier 1 · 99%
security

security-security-site-missing-x-frame-options-header-4d4da3fa

Tier 1 · 99%
security

security-security-site-missing-hsts-strict-transport-security-header-39631536

Tier 1 · 99%
security

security-security-site-missing-content-security-policy-header-723cd178

Tier 1 · 99%

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics