vendor-management
Pass
Audited by Gen Agent Trust Hub on Mar 23, 2026
Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill uses a direct interpolation of user-supplied input via the
$ARGUMENTSplaceholder in the prompt. An attacker could provide malicious arguments to override the agent's instructions or behavior. - [DATA_EXFILTRATION]: The skill attempts to access and modify a specific file path in the user's home directory:
~/.claude/projects/skill-telemetry.md. Writing execution metadata to hidden configuration directories is a sensitive operation that can be used to track user activity or maintain persistence. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). It is instructed to 'Read the actual codebase' and ingest 'vendor-related data structures' which may contain attacker-controlled content.
- Ingestion points: Reads files and data structures across the codebase (SKILL.md).
- Boundary markers: None identified. The instructions do not include delimiters or warnings to ignore embedded instructions in the analyzed data.
- Capability inventory: The skill has file system write capabilities (creating reports in
docs/and modifying telemetry in~/.claude/). - Sanitization: No evidence of input sanitization or validation of the ingested codebase/data is present.
- [DATA_EXPOSURE]: The skill specifically targets highly sensitive information including 'banking, tax status, certifications, insurance' and 'account manager' contacts. While the analysis is intended to be local, the processing of this data increases the risk surface for exposure.
Audit Metadata