How we test, in full daylight.
The five-stage pipeline behind every engagement, the standards we map to, and the decision tree that says when our AI keeps going and when a human takes over. Nothing about how we work is a trade secret. The advantage is in the people doing it.
Five stages. Each owned by whoever's best at it.
AI handles scale, speed, and the work humans were never going to enjoy. Senior pentesters handle scoping, direction, and the business-logic calls that make a finding real.
Discover · AI
Endpoint mapping, auth flow enumeration, token-type fingerprinting, exposed-surface inventory. The AI does this the way a real adversary would, day one of an engagement, without a sitemap.
Continuous AI pentest · AI
A hive of offensive agents tests every endpoint via direct API calls, chains exploits across services, and probes for the OWASP and CWE classes below. Runs 24/7, not just at quarter-end.
Human + AI deep dive · AI + Human
A senior offensive engineer directs the AI into business-logic territory, reproduces exploits end-to-end, validates the AI's flagged findings, and adds the chains the AI couldn't have reasoned about alone.
Compliance-ready report · AI
Findings auto-compile into a report mapped to GDPR Art. 32, SOC 2 CC6.1, ISO 27001, PCI-DSS, and HIPAA, with reproduction steps, blast radius, and remediation per finding. Delivered in days.
Retest on demand · AI
Ship a fix. Hit retest. Re-runs the exact chain that broke you, against the patched build. Verifies the fix or tells you what's still exposed. No new SoW, no new invoice.
Safe in production · Human + AI
Guardrails defined by our humans every engagement. We exploit to prove impact, not to cause it. No destructive actions, no data exfiltration, no service disruption. Scope agreed before we start.
When AI keeps going, and when a human takes the keyboard.
Most platforms wave their hands at "human-in-the-loop." Here's the actual decision tree.
AI runs alone
Endpoint discovery, auth flow mapping, known-class probing (OWASP Top 10 templates), CVE matching, and the first pass of exploit-chain hypotheses. The AI also writes the report draft and runs the retest after a fix is shipped.
- Discovery
- Known-class probing
- CVE matching
- Retest execution
AI flags, human decides
The AI surfaces a candidate finding that requires judgment. A senior reviewer confirms, refutes, or routes it deeper. The reviewer's note travels with the finding into the report. Every critical and high finding goes through this gate.
- Severity calibration
- Exploit reproduction
- False-positive culling
- Reviewer annotation
Human takes over
Business logic that requires understanding what the app is *for*. Multi-step workflows that compose three or more chained primitives. Cross-tenant isolation breaks. Anything involving real money, real identity, or destructive impact. Anything where the AI says "I'm not sure."
- Business logic
- Tenant isolation
- Payment / identity
- Multi-step chains
Human-only
Scoping calls. Final report sign-off. Any action that would have side effects on production. Any decision about whether to attempt a chain that the customer hasn't pre-authorized. The senior reviewer's name and credentials are on every report.
- Scoping
- Final sign-off
- Production-impact decisions
- Scope expansion
Mapped to the frameworks your auditor already accepts.
Every finding ships with the standard references your compliance team needs. Not after the fact. Not as a bolt-on PDF. In the report.
OWASP
Full coverage of OWASP Top 10 (Web), OWASP API Security Top 10, OWASP Mobile Top 10, and OWASP Top 10 for LLM Applications. Each finding tagged with its OWASP class. ASVS mapping available on request.
- Web Top 10
- API Top 10
- Mobile Top 10
- LLM Top 10
NIST
Test methodology aligned with NIST SP 800-115 (Technical Guide to Information Security Testing). Findings map to relevant NIST 800-53 controls where customers need that mapping for federal or federal-adjacent compliance.
- SP 800-115
- SP 800-53
- CSF 2.0 mapping
- Federal-friendly
CIS
For cloud config engagements, we run full CIS Benchmark coverage for AWS, GCP, and Azure. Each CIS control is checked against your account, with every finding tied to its real-world exploit path, not just its checkbox status.
- CIS AWS Foundations
- CIS GCP Foundations
- CIS Azure Foundations
- Exploit-path mapping
OWASP APTS
The new OWASP Autonomous Penetration Testing Standard defines eight governance domains for AI-driven testing in production. Our platform is designed against all eight. Trust page has the per-domain detail.
- Scope enforcement
- Safety controls
- Human oversight
- Auditability
Compliance frameworks
Findings auto-map to the control families your auditors actually cite, with citation per finding. The same report passes a SOC 2 Type II audit, an ISO 27001 surveillance audit, or a PCI-DSS attestation without rework.
- SOC 2 CC6.1, CC7.1
- ISO 27001 A.12.6.1, A.14.2.8
- GDPR Art. 32
- PCI-DSS 11.4, HIPAA §164.308
Our own research
The standards are the floor. The ceiling is what our team has found across two decades of offensive work: CVE-grade findings, exploitation primitives that haven't made it to the public corpus, and the business-logic patterns that only show up after you've broken a hundred apps in the same vertical.
- CVE corpus
- Internal exploit lib
- Vertical playbooks
- Published research
The classes we run, by hand and by agent.
Who can do what, and how they get in
IDOR and BOLA, broken authentication, JWT flaws (alg confusion, weak keys, key confusion), OAuth and SSO misconfiguration, privilege escalation chains, session fixation, session hijacking, and the SAML edge cases nobody documents.
Inputs that get turned into code
SQL injection (all flavors), SSRF and blind SSRF, command and OS injection, server-side template injection, XXE and unsafe deserialization, prototype pollution, and full RCE chains that compose three or more primitives.
Flaws scanners can't understand
Race conditions, workflow and state-machine abuse, payment and pricing tampering, multi-step logic bypass, rate-limit and quota abuse, tenant isolation gaps, and the "if you call these three endpoints in this order" bugs that need a human.
The surface your app really is
BFLA and object-property auth flaws, GraphQL query abuse, mass assignment, insecure mobile storage and IPC, binary-level reverse engineering, certificate and pinning issues, and the OAuth deep-link attacks that ship in every other Android app.
Misconfigs that compose into compromise
Over-privileged IAM roles, wildcard policies, cross-account trust gaps, public storage and snapshots, permissive security groups, unrestricted egress, KMS grant abuse, CloudTrail blind spots, and the CIS controls that map to real exploit paths.
The attack surface AI just added
Direct and indirect prompt injection, jailbreak chains, RAG context leakage, tool-call abuse and agent manipulation, guardrail bypass, training-data inference, and the auth gaps that show up when an LLM workflow can read someone else's session.
Want the full methodology doc? It comes with access.
Every customer gets the long-form methodology PDF, including the exploit-chain decision tree and the per-vertical playbooks our team uses.
Request access