📈 We Ran Claude on Our Codebase. Lorikeet Still Found Two Highs.

Here's a thing that's happening in 2026: engineering teams are running AI security audits before they even think about hiring a pentester. Claude reads the whole repo, finds the XSS, flags the SQLi, catches the weak crypto. It's genuinely good at this. Flowtriq did exactly this — ran a thorough Claude-assisted code review, took it seriously, fixed everything, shipped the patches.

Then Lorikeet Security came in and found five more findings anyway.

Not because the AI audit failed. Because the AI audit was done.

📈 The Part AI Can't See

All five remaining findings had one thing in common — none of them lived in source code. They lived in the running system. The deployed infrastructure. The stuff that only exists when the app is actually on and breathing.

Two High-severity session management issues. One endpoint had no rate limiting whatsoever — you only discover that by actually hammering it. Another had anti-forgery token validation that looked fine in code but fell apart under specific runtime conditions. You find that by replaying requests with mangled tokens and watching what the server actually does. No LLM can do that from a git clone.

A Medium around TLS. The production listener was still negotiating a deprecated protocol version — inherited from an Ansible role nobody had touched in eighteen months. Completely invisible to static analysis. Completely visible to anyone who speaks TLS at the live server.

A Low for files left on the document root. An engineer dropped some operational artifacts there during an incident and forgot. Not in the repo. Not referenced anywhere. Just sitting there. A manual tester finds this in five minutes.

A Low for missing security headers. Present on the main app, missing on a subdomain, inconsistent across some edge-case routing. Came from reverse proxy config nobody had reviewed end-to-end in a while.

📈 Why This Is Actually Good News

The AI audit closing the obvious stuff — XSS, SQLi, template injection, weak crypto — meant Lorikeet's testers didn't spend their hours on that. They went straight to runtime. Infrastructure. The second attack surface that most teams haven't even thought about yet.

Both stages together gave Flowtriq more total coverage than either would have alone. The AI pass was a force multiplier on the pentest, not a replacement for it.

All five findings were closed within 48 hours. Lorikeet retested two weeks later. Everything done, no regressions.

"We came in thinking our AI audit had probably caught most of what mattered, and the report made us realize it had caught most of what mattered in the source tree — the runtime and infrastructure were a whole second surface area we hadn't actually tested." — Jacob M., Founder, Flowtriq

📈 The Prompt That Actually Matters

Not "did AI find the bugs." It did. The prompt is: what did AI structurally cannot see? Session state. Live TLS negotiation. Files on disk. Reverse proxy headers. Config that lives outside the repo entirely.

That's where the residual risk went. That's where it's going for everyone running AI-assisted dev in 2026.

Read the full case study →