The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/build_hints.py uses subprocess.run to execute an external command provided through the --llm-ideas-cmd command-line argument. While shlex.split is employed to parse the command string, the execution of arbitrary user-provided commands is a sensitive operation that could be exploited if the argument is influenced by an attacker.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting untrusted data from external files and passing it to an external process. Findings include:
Ingestion points: Data is read from market_summary.json, anomalies.json, and news_reactions.csv (or .json) via the read_market_summary, read_anomalies, and read_news_reactions functions in scripts/build_hints.py.
Boundary markers: There are no boundary markers or instructions to the downstream tool to ignore embedded instructions within the processed data.
Capability inventory: The script executes subprocesses via subprocess.run in the generate_llm_hints function.
Sanitization: There is no evidence of sanitization or filtering of the textual content before it is passed to the external command.

edge-hint-extractor