The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructions perform scans of sensitive user configuration files, including ~/.zshrc, ~/.zprofile, and files within the ~/.config directory. Accessing these files to check environment variables can lead to the unintended exposure of hardcoded API keys, secrets, or authentication tokens within the model's context.- [COMMAND_EXECUTION]: The auditing workflow involves multiple shell commands using tools like find, grep, lsof, and package managers. These commands are constructed using variables derived from external security advisories.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it incorporates untrusted data from security advisories directly into command arguments without validation.
Ingestion points: Security advisory text provided as input to the agent.
Boundary markers: The skill lacks delimiters or explicit instructions to distinguish between trusted system commands and untrusted external data.
Capability inventory: The agent has access to a variety of system tools (bash, find, grep, launchctl, etc.) and file writing capabilities within the user's Documents folder.
Sanitization: No sanitization or escaping mechanisms are defined for variables like <pkg>, <binary>, or <VAR> which are populated from untrusted advisory content.

cyber-audit