paper-figure
Warn
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill generates standalone Python scripts (e.g.,
gen_fig_training_curves.py) at runtime based on the contents ofPAPER_PLAN.mdand experiment data. These scripts are dynamically constructed and written to the filesystem. - [COMMAND_EXECUTION]: In Step 5, the skill uses a shell loop (
for script in gen_fig*.py; do python "$script"; done) to execute the generated Python scripts. This pattern of generating and then immediately executing code is a significant attack surface. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources including
PAPER_PLAN.mdand various data files (JSON/CSV). This content is interpolated into generated scripts and passed to a secondary model (gpt-6-astra) for review. - Ingestion points:
PAPER_PLAN.md,figures/*.json,figures/*.csv, and project root data files. - Boundary markers: No specific delimiters or boundary warnings are implemented to isolate external data from instructions.
- Capability inventory: The skill has the capability to write files to disk, execute Python scripts, and spawn additional agent instances.
- Sanitization: There is no evidence of sanitization or validation logic for the content read from external data files before it is used in code generation.
Audit Metadata