The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The scripts (auth_test.py, recon_scan.py, sql_injection_test.py, vuln_scan.py, web_app_test.py, xss_test.py) make frequent use of subprocess.run to call external security binaries like nmap, sqlmap, hydra, and gobuster. While list-based arguments are used to prevent shell injection, the skill provides a high-risk capability surface for an AI agent to perform unauthorized network attacks.
PROMPT_INJECTION (LOW): Several scripts, notably generate_report.py and web_app_test.py, ingest external data from tool outputs (e.g., ZAP JSON reports) and interpolate this content into markdown reports without sanitization, creating an indirect injection surface.
Ingestion points: ReportGenerator.load_findings (JSON) and WebAppTester.run_owasp_zap (via zap_report.json).
Boundary markers: None present; external data is processed directly.
Capability inventory: Extensive subprocess execution, file system writes, and network tool orchestration.
Sanitization: No escaping or validation is performed on the ingested vulnerability descriptions or findings before they are rendered into reports.
EXTERNAL_DOWNLOADS (LOW): The skill documentation and internal error handlers encourage the installation of multiple third-party tools from various sources (e.g., pip install sublist3r, apt install hydra, sqlmap.org), which increases the overall attack surface and trust requirements.
COMMAND_EXECUTION (MEDIUM): In vuln_scan.py, the run_openvas_scan function constructs an OMP XML command using f-strings inside a subprocess argument (f'X<create_task>...<target hosts="{target}"/>...'). A malicious target string could potentially perform XML injection to manipulate the OpenVAS task configuration.

penetration-tester