In June 2026, Microsoft Defender Security Research led by Mohammed Zaid discovered that MCP tool descriptions can be quietly changed to redirect an agent into collecting and sending out sensitive data while every individual action still appears authorized.
Similarly, a scan of 1,808 MCP servers found 66% had some security finding (AgentSeal, 2026).
MCPTox benchmark also discovered that tool poisoning succeeds against major agents over 60% of the time, up to 72% on some models.
Vetit is solving that problem by vetting any MCP server before it is used by an agent, and unlike MCP scanners that stop at scanning the tool's descriptions, Vetit takes a different approach,
- It extracts the text from the tool's descriptions and identify malicious intents.
- It calls a tool to confirm what it ACTUALLY does, not what it says it does, that is, is a tool that states read-only hint truly read only?
- In some servers not all tool might be compromised, so Vetit identifies which tools are and states them in its final findings.
- Finally, it creates a finding JSON document for admin review, and generates a Trueforge config to add the server to Trueforge, enabling trusted tools and disabling untrusted ones, and in cases where the findings are critical, it generates no config.
Who is it for?
Vetit is for builders building agentic tools on Trueforge and wants a tool to verify external MCP servers before adding them to their workflow,
Vetit can also be used to test an MCP server in Dev to validate its processes.
Vetit is also useful for security and threat auditors.
Vetit comes in two forms, Vetit agent and Vetit decoy MCP (an intentionally compromised MCP for testing Vetit).
Vetit can also audit itself, so you can run Vetit on Vetit to scan it.
Try it out on GitHub: https://github.com/codewithveek/vetit
Top comments (0)