exercise-assignment-dividend-risk
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external CSV files through a diagnostic script, whose output is then used by the agent to evaluate risk controls. \n- Ingestion points: The
scripts/exercise_assignment_dividend_risk_diagnostics.pyscript reads user-supplied CSV files viapandas.read_csv. \n- Boundary markers: There are no explicit delimiters or instructions provided to the agent to isolate the ingested data from its reasoning process. \n- Capability inventory: The skill can read local files, write JSON artifacts, and output summaries to the console. \n- Sanitization: Although the script coerces data values to numeric and datetime formats, the raw column names are included in the output JSON, allowing potential instruction injection through CSV headers.
Audit Metadata