research-ml-practice
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data from external files and caller inputs, which could be manipulated to influence the agent's behavior during the research process.\n
- Ingestion points: The skill uses content from
JOURNALfiles,EDA reports, and named concerns provided by the calling skill to determine the scope and modality of the research (SKILL.md).\n - Boundary markers: The instructions do not define clear boundary markers or delimiters (such as XML tags or specific block quotes) to isolate external data from the skill's operational logic.\n
- Capability inventory: The skill is empowered to perform wide-net web searches using external engines and write files to the
scratch/research/directory (SKILL.md,references/search.md).\n - Sanitization: The skill mitigates risks by mandating an 'abstraction' step where specific dataset names and identifiers are removed from queries before they are sent to external services, effectively reducing the risk of accidental data exfiltration or targeted adversarial manipulation via search queries.
Audit Metadata