The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/auto_detect_candidates.py implements a --llm-ideas-cmd argument that allows the execution of arbitrary external commands via subprocess.run. This provides a significant surface for shell command execution if the input command string is influenced by an attacker.\n- [COMMAND_EXECUTION]: The script scripts/validate_candidate.py dynamically constructs a Python code snippet and executes it through a subprocess using uv run python -c. This runtime script generation and execution is a high-risk operation that could be abused if the interpolation of variables into the snippet is bypassed or manipulated.\n- [PROMPT_INJECTION]: The skill ingests free-form text from 'hints' files and external process outputs to score and prioritize market candidates. Malicious content within these inputs can manipulate the agent's research prioritization and logic.\n
Ingestion points: The --hints YAML input and the output of the command provided to --llm-ideas-cmd in scripts/auto_detect_candidates.py.\n
Boundary markers: None. Content is parsed as structured data but used directly in logic scoring without isolation.\n
Capability inventory: Subprocess execution (subprocess.run), local file creation and modification (write_outputs, export_candidate).\n
Sanitization: The skill performs basic alphanumeric sanitization for identifiers using regex, but lacks validation for the instructional content of hints processed during candidate scoring.

edge-candidate-agent