On this page
Threat Model
MCP Audit focuses on risk before an agent connects to a server or invokes a tool.
Covered threats include malicious or compromised servers, poisoned tool descriptions, unsafe command launchers, unauthenticated remote transports, secret leakage, prompt injection, data exfiltration, overly broad filesystem or network access, excessive agent authority, and unsafe combinations of individually acceptable tools.
Trust Questions
- Who publishes and maintains the server?
- What process, filesystem, network, and credential access can it obtain?
- How is a remote endpoint authenticated and encrypted?
- Can tool descriptions influence an agent beyond their legitimate purpose?
- Which actions require human approval?
- What evidence will remain after a decision or invocation?
A clean scan is one input to a decision; it is not a guarantee that a server or its dependencies are trustworthy.