dse-loop

Warn

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill instructs the agent to dynamically create and execute scripts (e.g., dse_results/parse_result.py) to extract metrics from program outputs. It also mandates discarding and reimplementing these scripts if they fail, which involves repeated generation and execution of code based on external file content.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources including user arguments, project source code (to infer parameter ranges), and raw program outputs. There are no explicit boundary markers or sanitization steps mentioned to prevent the agent from following instructions embedded in the source code or logs it analyzes during the exploration loop.
  • Ingestion points: User-provided $ARGUMENTS, project source code files, and program stdout/stderr logs.
  • Boundary markers: None specified; the agent is told to "Read the source code" and "Parse metrics programmatically" without delimiters.
  • Capability inventory: File read/write access and execution of shell commands, scripts, and build tools (gcc, yosys, etc.).
  • Sanitization: None described for the parsed content before it is used to influence the agent's logic.
  • [COMMAND_EXECUTION]: The exploration loop involves executing arbitrary commands provided in the Program argument or inferred from the project (e.g., make, python, gcc). This allows for the execution of arbitrary software within the environment as part of the tuning process.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 10, 2026, 09:36 AM