MCP Protocol Security

AegisGate provides comprehensive security enforcement for the Model Context Protocol (MCP), the standard protocol for AI model-tool interaction.

Threat Landscape

MCP enables AI models to invoke external tools, retrieve context, and interact with APIs. This creates several attack vectors:

  • Prompt injection via tool results: Malicious tool responses can inject instructions into the AI’s context window
  • Tool poisoning: Adversarial tool definitions that manipulate model behavior
  • Cross-server request smuggling: Attacks that tunnel through MCP server-to-server communication
  • Token limit abuse: Tools that return oversized responses to exhaust context windows

AegisGate Enforcement

LayerProtectionDetection Rate
Request screeningPrompt injection detection100% (adversarial suite)
Tool validationSchema enforcement, parameter sanitization42 attack patterns
Response filteringOutput classification, PII redaction83.1% coverage
Rate limitingPer-server, per-tool, per-clientConfigurable
Audit loggingFull MCP session recordingATLAS-mapped

Configuration

mcp:
  enabled: true
  screening:
    prompt_injection: block
    tool_poisoning: block
    token_abuse: warn
  response:
    filter_pii: true
    max_response_tokens: 4096

For general security architecture, see Security Overview.