vector-conflict-check
Pass
Audited by Gen Agent Trust Hub on Apr 29, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface when performing semantic conflict checks.
- Ingestion points: The skill retrieves vector excerpts and metadata from an external Pinecone index in steps 2 and 4 and provides this untrusted data to the model for analysis.
- Boundary markers: No explicit delimiters or instructions are used to segregate retrieved content from the model's instructions, creating a risk that malicious data in the vector index could influence the agent's behavior.
- Capability inventory: The skill possesses Write access for reconciliation reports, restricted Bash access (sha256sum, test, date, mkdir), and Pinecone MCP tools.
- Sanitization: No sanitization, escaping, or validation of the retrieved vector content is performed before interpolation into the model's context.
- [DATA_EXFILTRATION]: The skill reads sensitive local files such as configuration files, ground-truth assertions, and CRM records (crm/companies.md) to facilitate the synchronization process. This data access is consistent with the skill's purpose, and no unauthorized transmission to external domains was identified.
- [COMMAND_EXECUTION]: The skill utilizes the Bash tool for environment checks and file integrity verification. Access is restricted via frontmatter configuration to a specific subset of safe utilities including sha256sum, test, date, and mkdir.
Audit Metadata