The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill contains multiple instances of curl | bash and wget | sh strings within agent_guard.py (inside the demo_clinejection_detection function), SKILL.md (documentation examples), and test_agent_guard.py (unit tests). These are used as sample inputs to verify that the tool correctly identifies remote code execution attempts and are not executed as actual shell commands.
[PROMPT_INJECTION]: The core logic in agent_guard.py includes a list of injection patterns such as "ignore previous instructions" and "forget everything." These are utilized for string matching to protect the agent and do not represent an attempt to bypass safety guidelines.
[COMMAND_EXECUTION]: Destructive command strings like rm -rf / and sudo are present in the execution_patterns list in agent_guard.py and within test_agent_guard.py. These serve as validation data for the detection engine's sanitization capabilities.
[EXTERNAL_DOWNLOADS]: The documentation and test suite reference external URLs like evil.com and raw.githubusercontent.com to demonstrate how the tool flags suspicious network operations. No actual unauthorized network requests are made by the core analysis engine, which relies solely on the Python standard library.

agent-guard