The Agent Skills Directory

[COMMAND_EXECUTION]: The platform uses subprocess.run and os.system for critical operations, including cloning git repositories, managing Python virtual environments via uv, and executing entrypoint scripts for installed skills.
[REMOTE_CODE_EXECUTION]: The test suite includes several samples of remote script execution patterns, such as piping curl output from external domains directly into bash. The skill's primary setup instructions also encourage running a remote shell script from astral.sh to install the uv tool.
[DATA_EXFILTRATION]: The gauntlet test suite includes simulated attacks that read sensitive local files, such as SSH private keys and AWS credentials, and transmit them to external example servers.
[OBFUSCATION]: Malicious code patterns in the test suite are hidden using techniques such as Base64 encoding and hex-encoded character sequences to validate the detection logic of the integrated security scanner.
[CREDENTIALS_UNSAFE]: The security scanner's audit logs and test datasets contain various examples of hardcoded secrets, including AWS keys and GitHub tokens, used for testing credential detection.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of external code packages from GitHub and the Decision Hub registry. It also references the official uv installer from astral.sh.
[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the AI agent to automatically read and load the system prompts (SKILL.md) of newly installed skills into the current conversation, creating a potential path for untrusted third-party content to influence agent behavior.

my-skill