AI Agents for Authorized Penetration Testing: Lab

Abstract

A self-contained educational module on using AI agents safely within an authorized penetration test. It pairs a conceptual manuscript with a hands-on, fully synthetic and offline lab in which students turn authorization documents into a machine-enforced policy, adjudicate an AI agent’s actions through a policy decision point (a “policy gateway”), detect prompt-injection and false-authorization traps, audit an agent’s action log, and review an AI-written draft report — separating sound findings from unsupported ones. The package contains: (1) a manuscript presenting a human-governed workflow, an agent-specific threat model mapped to OWASP, MITRE, and NIST guidance, a propose–decide–act reference architecture, an evidence-and-provenance model, and an evaluation protocol; (2) a student lab guide (Parts A–G) with an assessment rubric; (3) an instructor guide and answer key; (4) a class introduction slide deck and a one-page session syllabus; and (5) a runnable, standard-library-only Python lab kit — an 18-item synthetic evidence packet, a policy gateway, a 30-case test suite, audit-metrics and finding-linter tools, and an autograder. All scenario data is fictional: hostnames use the reserved .example TLD and all IP addresses are from the RFC 5737 documentation ranges. The lab runs without network access and contains no exploit code; it teaches the governance and reasoning behind AI-assisted testing rather than offensive techniques. Primary references: NIST SP 800-115; OWASP Top 10 for LLM Applications and for Agentic Applications (2026); MITRE ATT&CK and ATLAS; NIST AI RMF (AI 100-1) and its Generative AI Profile (AI 600-1).