Tous les articles

Prevention Under Pressure: Why most AI governance is "detection wearing prevention clothes"

John Sapp Field CISO + 1 other

AI is changing both sides of the security equation. Companies are using it to build faster, and attackers are using it to hit harder. Security leaders are being pushed to rethink what prevention actually means.

That's the idea behind Prevention Under Pressure, a new Chainguard video series of short, candid conversations with CISOs about how AI-driven attacks are reshaping their threat models. Host John Sapp sits down with peers who are working out the answers in real time, and he doesn't pretend anyone has it all figured out yet.

The first guest is Kayla Williams, CISO and principal at Williams Rose AI Cyber Advisory and the creator of the GRC CISO framework. Here are a few highlights from the conversation.

The attack surface outran the program

Asked when she realized her existing approach to security wasn't going to hold up, Kayla pushed back on the framing a little. The program wasn't failing. Her controls were solid on paper and in practice, built to satisfy regulated clients at a global SaaS company. The trouble was speed. The attack surface was growing faster than any framework, tool, or team skill set could follow.

"Once AI hit the picture, it just exploded," she said. "Infrastructure was being spun up and spun down before we even knew it existed."

Her conclusion was that bolting AI onto inherited processes wouldn't work. Governance had to be rebuilt from scratch.

Phishing emails and a deepfake CEO

Kayla pointed to two moments that changed how she thought about AI-enabled threats. The first will be familiar to anyone who has run a security awareness program: phishing emails stopped looking like phishing. The misspellings and obvious tells disappeared. "It's clean, it's pristine, and it gets through your filters."

The second was more unsettling. At the end of a virtual conference keynote, the CEO walked onstage in a different outfit and revealed that the person who had just delivered the talk was an AI-generated version of them. Kayla had met this executive in person and never suspected a thing. "I knew we were in for a rude awakening," she said.

Before and after: Shrinking the blast radius

John asked Kayla to describe the day a major attack goes public, before and after adopting secure-by-default software like Chainguard.

Before, she said, the first few hours go to figuring out what happened, because detection tools tell you that something happened but not what. Then come the legal, communications, and incident response teams, and everyone on the call walks away with a slightly different version of the truth. She remembered her first incident after building a SOC: hours on an open bridge line with no clear data flows and no visibility into upstream and downstream dependencies.

With prevention built in at the source, the day starts with a much smaller blast radius. "The vulnerable component either wasn't in your environment or it never got the chance to be exploited."

John described the shift this way: the industry has spent two decades trying to shift left, and now there's a chance to start left.

"Detection wearing prevention clothes"

John asked what in a typical security stack is truly preventative. Kayla's answer was risk tiering. She means deciding, before anyone touches a tool, what is allowed and what controls it needs.

At Williams Rose, she and co-founder Olivia Rose use a model called TIDE: Tier, Intake, Determine controls, Enact decision. It exists because, in Kayla's words, "most AI governance programs are detection wearing prevention clothes. They monitor usage after the fact instead of deciding what's actually allowed before anyone touches the tool."

The model also lets teams move fast without having to choose between speed and control. Low-risk use cases inside the company's risk appetite get approved automatically. High-risk ones, usually anything involving sensitive data, get real scrutiny. When employees see that security only steps in when it matters, they're more willing to be open about what they're actually using.

Guardrails, alignment, and budget

Kayla also covered the guardrails she considers non-negotiable. Data classification comes first, and it has to include proprietary models and formulas, not just PII. Identity and access management comes right after. She raised the risk of "abandoned agents": an employee builds an AI agent for a critical process, then leaves, and nobody else knew it existed.

On budget, her advice was practical: before buying anything new, audit what you already own. The capability you need may be a feature you just haven't turned on. Then spend on the gaps that remain. That way you can show leadership you've done your homework.

Watch the full conversation

There's more in the full episode, including Kayla's thoughts on getting legal, HR, finance, and engineering aligned, John's "four generations of security leaders," and the capability she thinks most CISOs will need within the next 12 months.

Watch Episode 1 of Prevention Under Pressure with Kayla Williams.

Share this article

Commencez à développer avec Chainguard