AI-generated software is forcing yet another platform rethink

At a panel on security this summer at PlatformCon London, someone asked a room full of platform engineers to raise their hands if their teams were actively using AI to write and review code. Nearly everyone did. Then came the follow-up: keep your hand up if you feel your security systems can actually handle that. All but two hands dropped. That gap, between adoption and readiness, is the real story here, not the AI itself.

Liz Rice, chief open source officer at Isovalent, put it bluntly to The New Stack: finding vulnerabilities is "almost wasted work now" because AI systems behave non-deterministically. You can't predict what a model will do next, so the old idea of hunting bugs after the fact starts to feel outdated. Joe Baguley, CTO of EMEA at Broadcom, took it further during the panel, arguing there's no such thing as a zero-day anymore and that software bills of materials, and increasingly AI bills of materials covering model weights, training data, and third-party APIs, are about to become essential just to know what you've actually built.

Baguley also raised a scenario that sounds almost comic until you sit with it: seventeen different HR staffers each building their own AI agent to handle some sliver of their job. He calls it a Frankenstein's monster problem, and Rice's worry is sharper still, since HR agents built by non-technical staff are sitting on sensitive employee data with security nobody vetted. Nigel Douglas, head of developer relations at Cloudsmith, described the natural next step: someone builds a local tool that replicates paid software like Asana, shares it with less technical teammates, and suddenly the platform's user base includes people who were never meant to touch infrastructure like this.

The scale of the patching problem underscores why panelists kept circling back to speed. In 2025, a record 48,185 CVEs were published, with a fifth rated critical or high severity, and the average time to patch sat at 55 days according to EdgeScan's Vulnerabilities Statistics Report. AI models can now find and chain exploits automatically, so a 55-day window looks less like a delay and more like an open door. Rice's answer is runtime shields, often built on eBPF, that can block privilege escalation or flag odd network behavior while a proper patch is still being written. They're not a replacement for patching, just a way to survive the gap.

Baguley's closing point cuts to the structural question: this isn't only about controlling what AI agents do, it's about rebuilding platforms from the ground up because the security playbook from twenty years ago doesn't fit anymore. He calls for platforms that are leaner, harder, and faster, patchable almost automatically, the way a phone updates without the user noticing. Whether that means opening internal developer platforms to everyone in a company or locking things down with more traditional, centralized IT control is still an open argument nobody on that panel resolved.