The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone source code from an external repository located at https://github.com/zakirkun/guardian-cli.git during the setup phase.
[REMOTE_CODE_EXECUTION]: Following the download, the skill executes the external code by running pip install -e . and python -m cli.main, allowing arbitrary code from the third-party repository to run on the host system.
[COMMAND_EXECUTION]: The skill's primary functionality involves executing various shell-based security tools (e.g., nmap, nuclei, httpx, gobuster) using Python's subprocess.run. It also allows for the creation of 'Custom Tools' which provide a direct interface for executing arbitrary shell commands.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingests untrusted data from external targets—such as HTTP headers, service banners, and technology fingerprints—and passes this data to specialized AI agents (Planner, Analyst) for processing. An attacker-controlled target could provide malicious payloads designed to manipulate the AI's reasoning or the subsequent commands executed by the framework.
Ingestion points: Data returned from tools like httpx, nuclei, and whatweb is processed by the AI agents in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the AI guidance examples.
Capability inventory: Extensive system access including shell command execution (subprocess.run), network operations (via integrated tools), and file system writes (report generation).
Sanitization: The provided implementation examples do not demonstrate sanitization or validation of tool outputs before they are interpolated into AI prompts.

guardian-ai-pentest-cli