Model Context Protocol servers expose tools that LLM agents can invoke. Without proper authentication and monitoring, malicious agents can abuse these tools for data exfiltration, unauthorized actions, or resource exhaustion.
Discovery Vectors
Agents discover MCP servers through multiple channels:
/.well-known/mcp.json — standard MCP discovery endpoint
CounterAgent's JA4 fingerprinting can identify which TLS library an MCP client uses, even if it presents valid authentication tokens. This helps detect:
Credential theft — legitimate tokens used from unexpected TLS stacks
Tool abuse — automated invocations exceeding normal patterns
Prompt injection relays — agents forwarding injected instructions to your tools