trial-work-design
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [SAFE]: The skill provides a structured method for designing trial work plans. It isolates all mathematical and cost-assessment logic into local deterministic Python scripts (scripts/plan_trial_work.py), reducing the risk of prompt-based manipulation of financial data.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied business problems and references external sites for service conditions.\n
- Ingestion points: Business problem notes, budget data, and external documentation URLs defined in references/source-checks.md.\n
- Boundary markers: The skill uses scripts/plan_trial_work.py as a logic boundary for all data calculations. Explicit instructions in references/source-checks.md command the agent to ignore any directives found within external web pages.\n
- Capability inventory: Capabilities are limited to data analysis and plan generation; no network write or arbitrary command execution tools are utilized.\n
- Sanitization: Instructions mandate that the agent prioritize user-defined constraints and treat external content strictly as passive data, ignoring any embedded instructions.\n- [DATA_EXFILTRATION]: The skill includes comprehensive privacy safeguards. Instructions explicitly direct the agent to minimize PII usage, utilize pseudonyms for candidates, and store sensitive planning data in restricted local directories. Analysis of the included scripts confirms there is no logic for transmitting data to external servers.
Audit Metadata