barry-translation-training
Warn
Audited by Gen Agent Trust Hub on Jun 14, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill instructions direct the agent to read sensitive local file paths within the user's home directory. It attempts to locate and read a methodology index (
wiki/_index.json) across several possible paths including~/.agents/skills/,~/.claude/skills/, and~/.barry-english/methodology. It also specifically accesses~/.barry-english/profile.mdto retrieve user-specific metadata such as their CEFR level and profile information. - [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it is designed to ingest and process arbitrary text provided by the user for translation exercises.
- Ingestion points: User-provided translation materials and workplace content.
- Boundary markers: Absent; there are no instructions for the agent to distinguish between the content to be translated and potential instructions hidden within that content.
- Capability inventory: The skill maintains a complex logic flow that interprets and responds to the provided text across multiple steps (Filtering, Deverbalization, etc.).
- Sanitization: No evidence of input validation, escaping, or filtering for instructions embedded in the translation source.
Audit Metadata