skill-forge-eval
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external skill descriptions and user-defined evaluation sets to drive its multi-agent workflow, creating a surface for indirect prompt injection. \n
- Ingestion points: Ingests natural language instructions from external
SKILL.mdfiles during auto-generation (Step 1b) and fromevals/evals.jsonor user-specified files (Step 1). \n - Boundary markers: The prompt templates for sub-agents lack explicit delimiters or instructions to ignore potential commands embedded within the evaluated content. \n
- Capability inventory: The skill spawns sub-agents (
executor,grader,analyzer), executes local Python scripts, and performs extensive filesystem write operations. \n - Sanitization: There is no evidence of sanitization, escaping, or validation for the content ingested from external skills or user prompts. \n- [COMMAND_EXECUTION]: The skill automates core functionality by invoking local Python scripts. \n
- Evidence: Uses subprocess calls to execute
scripts/generate_eval_set.pyfor bootstrapping test sets andscripts/aggregate_benchmark.pyfor data processing. These scripts are internal to the skill's package.
Audit Metadata