Security Overview

mcp-bridge includes a built-in security layer for tool exposure and tool results:

• Trust Levels: per-server result handling (trusted, untrusted, sanitize)
• Tool Filter: allow/deny tool visibility and execution per server
• Result Limits: global/per-server output size caps with truncation metadata

This layer provides baseline protections for common threats:

• Known prompt-injection string patterns
• Oversized tool responses
• Unauthorized tool use through allow/deny controls

Limitations

The built-in layer is intentionally simple and does not cover:

• Regex-based sanitization only (pattern matching, not deep semantic analysis)
• Unicode obfuscation / homoglyph attack protection
• Content-level PII detection
• Sophisticated multi-step injection chains

For production workloads with stricter requirements, add an external filtering layer (for example guardrails and/or PII redaction) between mcp-bridge and your application.