The Agent Skills Directory

[COMMAND_EXECUTION]: The Level 0 template in 'scripts/init_agent.py' and the 'references/minimal-agent.py' implementation provide a 'bash' tool that executes arbitrary shell commands via 'subprocess.run(shell=True)' without any input validation or safety filtering.
[COMMAND_EXECUTION]: The Level 1 template and 'references/tool-templates.py' implement a basic string-matching blacklist for 'dangerous' commands. This security control is insufficient as it can be easily bypassed by using command substitution, alternative shell syntax, or environment variables.
[DATA_EXFILTRATION]: The file-handling tools in 'references/minimal-agent.py' lack directory traversal protections. An agent could be manipulated into reading or writing sensitive files outside the intended workspace by using '../' sequences in file paths.
[REMOTE_CODE_EXECUTION]: The unconstrained 'bash' tool included in multiple templates allows the system to download and execute arbitrary external scripts (e.g., via 'curl | bash'), creating a direct path for remote code execution from untrusted inputs.

agent-builder