case-study-plan
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple local files, creating a surface for potential indirect prompt injection attacks where malicious instructions in the source files could influence the agent's behavior.
- Ingestion points: Reads data from
~/.claude-marketing/brands/_active-brand.json,~/.claude-marketing/brands/{slug}/profile.json,~/.claude-marketing/brands/{slug}/guidelines/_manifest.json, and files within~/.claude-marketing/sops/(as specified in the Process section). - Boundary markers: The instructions lack specific delimiters or "ignore instructions" warnings to prevent the agent from obeying commands embedded within the loaded profile or SOP files.
- Capability inventory: The skill uses the loaded data to generate narrative frameworks, interview questions, data visualization plans, and executive summaries. It does not possess network communication or shell execution capabilities.
- Sanitization: No evidence of sanitization, filtering, or validation of the content read from the local files was found in the skill's instructions.
Audit Metadata