The Agent Skills Directory

[DYNAMIC_EXECUTION]: The references/subagent-experiment-supervision.md file defines a pattern for automated experiment supervision where a monitoring loop checks logs for failure patterns and applies 'fixes' by modifying shell commands at runtime. This automated command modification logic can be abused to execute unauthorized commands if the log content is influenced by untrusted data.- [REMOTE_CODE_EXECUTION]: The supervision protocol includes a fix_pip_install function that extracts module names from ModuleNotFoundError patterns in logs and executes pip install. This allows for the dynamic installation of arbitrary remote packages if an attacker can trigger specific error messages containing malicious package names.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external CSV datasets and monitor execution logs. The automated 'self-healing' logic lacks explicit sanitization or boundary markers to prevent 'log poisoning,' where malicious data causes an error that triggers the automated package installation or command modification logic.- [COMMAND_EXECUTION]: Several components, including scripts/detect-compute.py and the supervision reference, use the subprocess module to execute system commands and manage external processes. While hardware probes are hardcoded, the general design encourages the agent to interact extensively with the shell and system environment.

data-scientist