Scanning a public skill with v0.5.16 produces a W008 high at risk 1.00, and the finding's own text says no secret was found.
Command:
uvx snyk-agent-scan@latest <repo>/skills/building-flutter-apps
Output:
Secret detected (high risk: 1.00). The pre-filtered potential matches include literal redaction markers of the form "**REDACTED_SECRET_*...**". Per the scanning rules a redaction marker of that exact form indicates a real credential was found and then stripped before scanning, and must be treated as a true secret (not a placeholder). I found no actual high-entropy secret values in the visible file contents, so there is no file/line to point to inside the provided skill files; the evidence comes from the separate <potential_matches> redaction markers.
The reasoning looks circular. The pre-filter redacts something, then the judge reads its own redaction marker as evidence that a credential was there.
What I checked in the source:
The string REDACTED does not appear anywhere in the skill. There are no secret-shaped values either, nothing matching ghp_, gho_, github_pat_, sk-, AKIA, xox, or a PEM header. The only hits for secret-like terms are GitHub Actions references such as ${{ secrets.WINDOWS_PUBLISHER_TOKEN }} in assets/windows-installer-workflow.yml, and prose telling readers to keep tokens out of source, for example "SENTRY_AUTH_TOKEN = build-only secret; never source + app config + runtime bundle" in references/error-reporting.md.
Repro is public: https://github.com/sgaabdu4/building-flutter-apps, skill at skills/building-flutter-apps, 41 instruction files and 8 assets.
Two reasons this matters. A high at risk 1.00 with no file or line isn't actionable by the author, and these results feed the public security panels on skills.sh.
Version 0.5.16, macOS, run via uvx.
Scanning a public skill with v0.5.16 produces a W008 high at risk 1.00, and the finding's own text says no secret was found.
Command:
Output:
The reasoning looks circular. The pre-filter redacts something, then the judge reads its own redaction marker as evidence that a credential was there.
What I checked in the source:
The string
REDACTEDdoes not appear anywhere in the skill. There are no secret-shaped values either, nothing matchingghp_,gho_,github_pat_,sk-,AKIA,xox, or a PEM header. The only hits for secret-like terms are GitHub Actions references such as${{ secrets.WINDOWS_PUBLISHER_TOKEN }}inassets/windows-installer-workflow.yml, and prose telling readers to keep tokens out of source, for example "SENTRY_AUTH_TOKEN= build-only secret; never source + app config + runtime bundle" inreferences/error-reporting.md.Repro is public: https://github.com/sgaabdu4/building-flutter-apps, skill at
skills/building-flutter-apps, 41 instruction files and 8 assets.Two reasons this matters. A high at risk 1.00 with no file or line isn't actionable by the author, and these results feed the public security panels on skills.sh.
Version 0.5.16, macOS, run via
uvx.