trainer-synthesize
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted external data (CSV, JSON, tables) to generate evaluation manifests and training datasets.
- Ingestion points: The skill accepts source material via csv_file, JSON inputs, and schema definitions in SKILL.md.
- Boundary markers: The instructions lack specific guidance on using structural delimiters or sanitization when interpolating this untrusted data into generated prompts, which could allow embedded instructions to influence the agent or downstream optimizers.
- Capability inventory: The skill uses scripts/run_synthesize.py for file reading and involves the agent writing files (evals/evals.json, train.jsonl).
- Sanitization: While the skill mandates independent verification and filtering of synthetic candidates, it does not explicitly require sanitization or escaping of the input source material to prevent injection attacks.
Audit Metadata