FOR AI AGENT DEVELOPERS

The skill package came from someone you have never met. Verify it before you install it.

Agent skills, MCP servers, plugins, and generated code arrive from marketplaces, repositories, and other people's models, and they run with your credentials. A malicious instruction or an undisclosed capability inside them does not announce itself. It reads like the rest of the code.

The risks concentrate where code meets trust: prompt-injection payloads embedded in instructions and data files, code paths that reach for credentials or tokens they have no reason to touch, functionality present in the package that the documentation never mentions. Reading the code helps and does not scale, and asking another model to review it inherits the blind spots of models like the one that wrote it.

The AI is the witness. The math is the judge. GauntletScore verifies each claim against primary sources, then computes the trust score with deterministic math. The agents gather the evidence; no model sits in the verdict. The same evidence always produces the same score.

What GauntletScore checks
Instruction and data files for prompt-injection payloads, so the package does not carry directives aimed at the agent that will run it.
Credential and secret access in the code paths, so the package touches only the access its stated purpose requires.
Hidden functionality against the package's own documentation, so what it does matches what it says it does.
Factual claims in documentation against the record, so the package's stated dependencies, sources, and references hold.

GauntletScore runs a dedicated analysis of the package's behavioral claims, testing whether the documented purpose, the declared permissions, and the actual code paths tell one consistent story. A package whose story does not hold counts heavily against the score.

What the first run looks like

Upload the package at gauntletscore.com. Minutes later you are reading the Gauntlet Report: a trust score with a 95% interval, every flagged finding with the evidence behind it, and a cryptographically signed, tamper-evident certificate you can keep with the dependency record. In a live production run, the engine examined nineteen claims in one fluent, credible document and returned two debunked against the court record, each with the source that contradicts it. Every voting agent independently recommended against proceeding.

LIMITS

It does not replace a security review for high-stakes deployments and it does not verify behavior it cannot observe in the package. A finding it cannot confirm is reported as unverifiable, not as malicious.

Where it fits
Before installation.

Run any third-party skill, server, or plugin before it gets your credentials.

Dependency review.

Keep the signed certificate with the dependency record as proof the package was checked at adoption.

Generated code.

Run code an AI produced for you before it ships into your agent.

There is a package in your queue right now that you were about to install on faith. Run it. Either it is what it says it is, or it is not, and you found out first.

Analyze a Document
Three free credits. No card. No call.