SkillSpector: fail your CI build on a risky agent skill

The Stack

Steps

1. 1

   Write the CI gate and prove it on a risky and a clean SARIF

   SkillSpector emits SARIF 2.1.0 (skillspector scan ./skill --format sarif > out.sarif). The CI gate reads that SARIF and fails the build on any finding at level error or security-severity >= 7.0. CI here proves the gate blocks a CRITICAL skill and passes a clean one; the scan that produces the SARIF needs the tool installed and is fenced.

   cat > risky.sarif <<'JSON'
   {
     "version": "2.1.0",
     "runs": [
       {
         "tool": { "driver": { "name": "SkillSpector", "informationUri": "https://github.com/NVIDIA/SkillSpector" } },
         "results": [
           {
             "ruleId": "data-exfiltration",
             "level": "error",
             "properties": { "security-severity": "9.1" },
             "message": { "text": "Skill posts file contents to an external host" }
           }
         ]
       }
     ]
   }
   JSON
   cat > clean.sarif <<'JSON'
   {
     "version": "2.1.0",
     "runs": [
       { "tool": { "driver": { "name": "SkillSpector" } }, "results": [] }
     ]
   }
   JSON
   node -e '
   const fs = require("fs");
   function gate(path) {
     const doc = JSON.parse(fs.readFileSync(path, "utf8"));
     const runs = doc.runs || [];
     let worst = 0, blocked = false;
     for (const run of runs) {
       for (const r of (run.results || [])) {
         const sev = parseFloat((r.properties || {})["security-severity"] || "0") || 0;
         if (sev > worst) worst = sev;
         if (r.level === "error" || sev >= 7.0) blocked = true;
       }
     }
     return { blocked: blocked, worst: worst };
   }
   function bad(m) { console.error("BAD: " + m); process.exit(1); }
   const risky = gate("risky.sarif");
   const clean = gate("clean.sarif");
   if (!risky.blocked) bad("gate let a CRITICAL skill through");
   if (clean.blocked) bad("gate failed a clean scan");
   console.log("gate OK: blocked the CRITICAL skill (severity " + risky.worst.toFixed(1) + "), passed the clean scan");
   '

2. 2

   Run the real scan in CI (the model step, not checked by CI here)

   Install the scanner (uv tool install git+https://github.com/NVIDIA/skillspector.git), then in your workflow run `skillspector scan ./path --format sarif > skillspector.sarif` on a directory, file, repo URL, or zip, and pipe it into the gate above. The static analysis plus optional LLM pass is fenced; only the gate logic is asserted here.