T01 · Skill Instruction Hijacking
- Location
scripts/content-ingest.sh:33- Finding
Untrusted reference content can hijack downstream Agent instructions
- Content
View full analysis
"$OUT/SKILL.md" <- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a transparent demo for turning one user-chosen file into a publishable skill package, but users should review copied content and path metadata before sharing the result.
Install only if you are comfortable with a manual converter that packages the full input file and its original path into a publishable skill. Before publishing generated output, inspect the copied reference file, remove sensitive content, sanitize local paths, and avoid ingesting untrusted documents without checking for prompt-injection text.
scripts/content-ingest.sh:33Untrusted reference content can hijack downstream Agent instructions
scripts/content-ingest.sh:35Unescaped input paths and filenames permit generated document injection
scripts/content-ingest.sh:50Publishable packages disclose local source paths
Referenced artifact was not completely inspected
- `SKILL.md` with name/description populated from the input filename and an auto-computed summary.
The skill explicitly packages user-supplied content by copying the source file into references/ and recording source_path metadata, but the user-facing description does not warn about this data propagation. This can unintentionally expose sensitive file contents or local filesystem paths when the resulting package is shared or published, especially because the workflow encourages publishing the output.