The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill accesses sensitive user configuration files to check for environment variables that could impact security posture.
Evidence: Searches shell initialization files (~/.bashrc, ~/.zshrc, ~/.profile, ~/.zprofile) for variable definitions.
Evidence: Inspects system-level environment settings via systemctl --user show-environment and launchctl getenv.
[EXTERNAL_DOWNLOADS]: The skill is designed to fetch security advisory content from external sources to ground its audit findings.
Evidence: The workflow requires the agent to fetch advisories from "authoritative sources" or use user-provided text before performing an audit.
[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted advisory text to guide its auditing logic, creating an attack surface for indirect instructions.
Ingestion points: Fetches advisories from remote sources or uses user-supplied advisory text (SKILL.md).
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat external advisory content as non-executable data.
Capability inventory: The skill has extensive command execution capabilities, including recursive filesystem searches and process inspection across multiple package ecosystems (npm, pip, cargo).
Sanitization: There are no instructions to sanitize or filter the content of the advisory before the agent interprets it to decide which checks to run.
[COMMAND_EXECUTION]: The skill executes a wide array of diagnostic commands to identify installed software and running services.
Evidence: Uses find, grep, and xargs to locate package locks and manifests across the home directory and workspace.
Evidence: Invokes package manager listing commands such as npm ls, pip list, cargo install --list, and brew list --versions.
Evidence: Inspects system state using pgrep, ss, and lsof.

cyber-audit