Every category that's shifted work into the developer's live coding session — testing, linting, even infrastructure cost-checks — has already had its "moved into the IDE" moment. Application security hasn't, and not for lack of trying. The two established ways to test for vulnerabilities were both wrong for that inner loop: reading the code is fast but blind to how an app actually behaves, and testing a deployed environment is accurate but needs a pipeline stage and usually a ticket. Neither fits inside the minute an AI coding agent takes between finishing a feature and moving to the next one.

This session covers what changed, and what happened when that loop got closed instead of just shortened. AI coding agents can now stand up the application they just built, in the same session, and test its actual behavior in seconds — which means find, fix, and verify can be one loop instead of three handoffs. We'll share what two months of that running in production looks like: 3,200 vulnerabilities found and fixed by nine-plus different coding agents, a 98% rate of staying fixed, and no human security review in the loop. Then we'll walk through real, anonymized cases — a SQL-injection-plus-session-hijack chain that's the textbook shape of a records breach, and a remote-code-execution flaw on a payment-processing server that needed no chaining at all to be dangerous.

If you're an engineering leader trying to figure out whether "AI fixes vulnerabilities automatically" is a real operating model or a demo, this is the mechanism and the data, not the pitch.

Key Takeaways:

• Why AppSec is the one shift-left category that hasn't moved into the coding session yet — and what specifically changed to make it possible now, versus reading code or scanning a deployed environment.

• What "find-fix-verify as one loop" produced in two months of production use: 3,200 vulnerabilities closed, a 98% stay-fixed rate, zero human security review in the loop — the mechanism and the proof, together.

• What these flaws would have enabled left in place, verified by exploitation rather than pattern-matching — anonymized cases including a chain that needed no chaining at all, plus what got thrown out for not clearing that bar.