The Agent Skills Directory

[DATA_EXFILTRATION]: The scripts/failure_triage.py script implements a file-reading function (_load_log) that accepts an arbitrary file path via the --log-file command-line argument. The script does not perform any path sanitization, directory traversal checks, or sandboxing to restrict access to a specific workspace. As a result, the tool can be used to read any file accessible by the system process, which could include sensitive configuration files, environment variables, or private keys if an attacker influences the path argument.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external simulation log files which are often controlled by external inputs or environments. The script extracts a 500-character excerpt from these logs and includes it in its output. Since this output is returned to the agent, it creates an attack surface where instructions embedded in log files could influence the agent's subsequent behavior. The skill lacks explicit sanitization or distinct boundary markers to separate this untrusted content from the tool's diagnostic output. The ingestion points are the --log-file and --log-text arguments in scripts/failure_triage.py, and the capabilities involved include Read and Bash access.
[COMMAND_EXECUTION]: The workflow involves executing a local Python script via Bash. While the script's logic is primarily focused on string matching, the reliance on external command execution for triage logic, combined with broad file-access permissions, is a relevant security consideration.

simulation-failure-triage