multi-copywrite-alchemy

Pass

Audited by Gen Agent Trust Hub on Sep 11, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPERSISTENCEEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from social media articles to distill writing styles, creating an inherent risk that embedded malicious instructions could influence the generated persona's behavior. \n
  • Ingestion points: Articles are fetched from WeChat and Douyin via user-supplied identifiers as documented in references/core_workflow.md.\n
  • Boundary markers: The process includes text cleaning and pre-processing, but lacks explicit filtering for prompt injection vectors within the source material.\n
  • Capability inventory: Distilled findings are written to new SKILL.md files which the agent can subsequently invoke.\n
  • Sanitization: While the skill statistical analysis reduces the likelihood of direct execution of injected commands, there is no formal prompt injection validation on input articles.\n- [PERSISTENCE]: The skill creates new sub-skills in the local filesystem (~/.workbuddy/skills/[author]-style/) based on the distillation results. This mechanism allows the agent to maintain and invoke the analyzed writing styles across future sessions.\n- [EXTERNAL_DOWNLOADS]: The skill makes authenticated network requests to the vendor's API (redfox.hk) to retrieve article content and metadata required for analysis. This is a legitimate functional requirement for the skill's automated collection features.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 11, 2026, 04:51 AM
Security Audit — agent-trust-hub — multi-copywrite-alchemy