dedupe-against-corpus
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests content from external Markdown files located in
corpus/**/*.mdto calculate similarity scores and detect duplicates. While this involves processing untrusted data, the operation is safe due to the mechanical nature of the analysis (SHA256 hashing and Jaccard similarity) and the restricted writing capability. - Ingestion points: The skill reads the body and frontmatter of all files in the corpus directory during Step 1, Step 3, and Step 4.
- Boundary markers: No specific delimiters are used when reading the file content for comparison.
- Capability inventory: The skill can modify existing files, but its access is strictly limited to appending an ID to the
related_seedsfrontmatter field. - Sanitization: Comparison logic is based on statistical and cryptographic algorithms rather than natural language interpretation, which significantly reduces the risk of instruction execution from the ingested data.
Audit Metadata