The Framework for Testing AI Attack Surfaces

Traditional penetration testing methodology was designed for a world where software was static, APIs were synchronous, and nothing acted autonomously. That world is gone.

APEX - the Agentic Penetration Exercise methodology - was built from first principles for the AI era. It systematically tests the attack vectors that emerge when software can read instructions, call tools, maintain memory, and take autonomous actions.

The Five Phases

PLAN - Every APEX engagement begins with a structured threat model. We identify your AI agent architecture, map trust boundaries between components, define the rules of engagement, and run automated OSINT to understand your AI stack’s external exposure. AI agents gather public intelligence in parallel; human researchers synthesize it into a focused attack plan.

SURFACE - We enumerate your complete AI attack surface: every agent, every tool connection, every API endpoint, every MCP server, every privilege scope. Most engineering teams discover agents and tool integrations they didn’t know existed. This phase produces the attack surface map that drives the exploitation phase.

EXPLOIT - Human senior researchers drive creative attack chaining while AI agents run automated prompt injection sweeps (Garak, PyRIT, PromptBench) in parallel. This combination covers attack surface that manual testing alone cannot enumerate - at the speed that automated tools alone cannot contextualize. Critical findings are reported within 48 hours of discovery.

PERSIST - We simulate long-term adversarial presence: persistent agent compromise, privilege escalation through agent tool chains, exfiltration path mapping. This phase answers the question that standard pentests don’t ask: if an adversary got in, how long could they stay, and what would they do?

REPORT - Every finding is documented with business impact, CVSS score, reproduction steps, and remediation guidance. The report maps findings to OWASP LLM Top 10, ISO 27001, SOC 2, and NIST AI RMF - structured for use as compliance and audit evidence.

Human-Led, AI-Augmented

APEX is not an automated scanner. AI agents in APEX automate the high-volume, systematic work - enumeration, fuzzing, injection sweeps. Human senior researchers drive creative attack chaining, business logic exploitation, and findings narrative.

This combination eliminates the false-positive noise that purely automated tools produce, while covering attack surface that purely manual testing cannot enumerate in a reasonable timeframe.

APEX and Your QA Cycle

APEX was designed to integrate with the engineering workflows your team already uses. Each phase maps directly to your sprint and QA cycle:

PLAN aligns with sprint planning and threat modeling sessions. We arrive having already reviewed your architecture - your sprint planning slot becomes a focused threat model review, not a discovery call.

SURFACE can integrate with your existing asset inventory. We map APEX asset discovery against your internal service catalog, CMDB, or infrastructure-as-code definitions. Gaps between what APEX finds and what your inventory shows are findings in themselves.

EXPLOIT runs in parallel with your QA sprint. While your QA engineers run functional and integration tests, APEX researchers run adversarial tests against the same build. Findings land in your issue tracker alongside your QA tickets.

PERSIST simulates real adversary persistence across deployment cycles. We test whether a compromise achieved in one sprint survives your next deployment - validating that your CI/CD pipeline does not re-introduce cleared vulnerabilities.

REPORT produces compliance evidence mapped to ISO 27001 / SOC 2 / NIST AI RMF. The APEX attestation letter and structured findings report give your compliance team audit-ready documentation without requiring a separate evidence-gathering exercise.