Disciplined scope. Autonomous depth. Human verification.

We never look at a system until its owner has named the exact domains or hosts and approved the review in writing. Choosing a plan or sending us a message is not permission. Before work starts, both sides sign a short written scope listing the target domains, the test logins you are providing, any off-limits workflows, and the testing window.

An outbound scope filter sits between our testing harness and the internet. It allows traffic only to the domains named in your signed scope. If your website links out to a third-party service (such as Stripe, Shopify, or Calendly) or to another system you did not include in the scope (such as an internal ERP or a legacy subdomain), our models never probe it.

Instead, we record those outgoing links in an 'Observed Outside of Scope (Not Tested)' section at the end of your report. That way, if your public site accidentally links to an internal ERP login or a forgotten staging server, you find out about it without us ever stepping outside our authorization.

General-purpose frontier models are built to be polite, all-purpose assistants. When people talk about 'fine-tuning' in AI, they usually mean collecting a giant dataset of past work and training a model on it. We don't do that.

We run our own open-weight models whose weights we modify directly for offensive and defensive security work, paired with testing logic and system prompts that our reviewers edit by hand whenever we uncover new attack patterns. We maintain no training dataset whatsoever, and we never train on client data, source responses, credentials, or findings.

Because we started working on AI-assisted security shortly after ChatGPT launched in November 2022, four years of real-world edge cases and false-positive filters are baked directly into our model weights and testing harness.

Depending on your plan, our models and reviewers examine your systems from the outside like an unauthenticated visitor and from the inside using test logins you share through our encrypted client portal:

• Public perimeter & hygiene: pages, forms, scripts, outdated CMS plugins and frameworks, forgotten staging hosts, and private documents (audit reports, CSV exports, backups) left at public URLs.
• Role & permission boundaries (IDOR / broken access control): signing in as Customer A, Customer B, Staff, and Admin to verify whether a lower-privilege account can view, edit, or export another customer's records or reach staff-only endpoints.
• Business logic & chained paths: testing how multi-step workflows (checkout, quoting, password resets, file sharing, account invites) behave when steps are skipped, repeated, or combined.
• Your own AI features (chatbots, RAG search, copilots & agents): testing whether Account A can prompt your AI assistant into quoting Account B's documents from a shared vector database; checking whether tool-calling agents can be tricked into running privileged backend actions ('confused deputy' flaws); testing indirect prompt injection through forms and uploads; and checking for exposed LLM API keys or unthrottled endpoints ('denial of wallet').

In 2026, nobody wants a human clicking 'Approve' on every single HTTP request—that would erase the speed and thoroughness that make AI testing valuable. Instead, we place human judgment at the start and finish of the engagement, and wrap the autonomous run in strict technical guardrails:

• Before the run (Human setup): A lead reviewer verifies your signed scope, configures the domain allowlist, sets up your test accounts in the portal, and marks any sensitive endpoints (like live billing or mass-email triggers) off limits.
• Rate & latency circuit breakers: Our harness enforces a strict request-per-second ceiling and monitors target response times and 5xx errors. If your server slows down, testing throttles or pauses automatically so a review never causes an outage.
• Safe write guardrails inside admin/staff logins: Discovery and read checks run autonomously. When testing state-changing requests (edits, uploads, deletions) inside an authenticated role, our harness restricts mutations to self-created canary records owned by the test accounts—never deleting real customer orders or altering live company settings.
• After the run (Human verification): Every candidate finding is checked by a human reviewer before it enters your report. If we cannot reproduce it with clear proof, we throw it out.

Some work is a different kind of engagement, and we never fold it in unannounced:

• Inboxes, phones, and people. We do not phish or socially engineer your staff.
• Offices, badges, and physical networks.
• Volumetric load or denial-of-service (DoS) testing designed to knock a system offline.
• Third-party vendor infrastructure (such as Stripe, AWS, HubSpot, or Calendly) that you do not own.
• Any domain or host that was not listed in the signed written scope.

You get a concise written report ordered by what to fix first. Every finding contains four parts: (1) What we found, in plain language; (2) Why it matters to your business; (3) The proof (the exact URL, test account, request/response evidence, and timestamp); and (4) What to change.

Each finding is also tagged with its industry-standard category—mapping to the OWASP Web Security Testing Guide (WSTG), OWASP API Security Top 10, OWASP Top 10 for LLM & GenAI Applications, and PTES. That mapping means your report is immediately accepted by enterprise customer procurement teams, cyber insurance underwriters, and SOC 2, ISO 27001, HIPAA, or PCI-DSS auditors.

After your team makes the fixes and we complete your complimentary re-scan, we issue two documents: the updated technical report marking each item Closed or Still Open with fresh proof, and a signed 1-page Letter of Attestation confirming the scope, methodology, dates, and remediation status. You can hand that 1-page letter directly to customers, insurers, and auditors without exposing your internal technical details.

Your team (or your agency) makes the fixes. We never sell remediation hours, which keeps our review independent for compliance and audit purposes. If you want outside hands to make the repairs, we are glad to introduce trusted development partners.

While your team works through the list, we answer questions in the portal or by email so fixes land cleanly on the first try. Within 45 days of the report (60 days on the Infrastructure plan), we run the complimentary re-scan of the same scope at no charge.

Once the re-scan and Letter of Attestation are delivered, the engagement is complete. You can delete your shared test logins in the client portal at any time, and we automatically purge any remaining stored credentials 14 days after the re-scan wraps up.