A $25 Tool Just Broke the Exploit Market

GPT-5.6 found a $500,000 WordPress RCE for $25. That’s not a party trick. That’s a structural shift in who gets to play offense.

The economics of vulnerability research have always been the moat. Exploit brokers like Zerodium or crowdsourced platforms like HackerOne exist because finding critical bugs is hard, expensive, and slow. It takes skilled humans thousands of hours. The price of a WordPress remote code execution reflects that scarcity — half a million dollars is what the market says it costs to discover something that dangerous. Until now, that scarcity was load-bearing for the entire ecosystem: defenders could take comfort knowing that only well-resourced actors (nation-states, sophisticated criminal groups, elite researchers) could produce weaponizable zero-days at scale.

That comfort is gone.

The SLC Cyber post is worth reading carefully. A researcher used GPT-5.6 — not a fine-tuned offensive model, not some state-sponsored jailbreak, just a current frontier model — plus $25 in API calls to find a genuine RCE in WordPress. Reproduce that process a thousand times across a thousand popular software packages and you have something that looks less like a security research paper and more like an industrialized attack factory. The barrier wasn’t just cost, it was expertise. AI collapses both simultaneously.

Everyone says this changes defense too. The opposite is closer to true in the short run, because offense and defense are not symmetric. Offense needs to find one hole; defense needs to close all of them. AI amplifies that asymmetry before it corrects it. Yes, the same tools can be used to fuzz your own code, scan your own stack, audit your own dependencies. But the attacker gets to choose where and when. The defender has to protect everywhere, always. AI makes the attacker faster. It makes the defender faster too — but faster at a game they were already losing at scale.

The UK’s AI Security Institute quietly flagged something relevant here. Their analysis of the gap between open and closed-weight models on cyber-relevant tasks found the gap is shrinking. That’s the second shoe. Right now GPT-5.6 found this bug. Six months from now, an open-weight model with comparable capability finds it — and that model is downloadable, forkable, runs locally, and leaves no API logs. The economics get worse before they get better. The marginal cost of a serious exploit approaches zero when the model that finds it costs nothing to run and nothing to trace.

The UK AISI data and this WordPress story together tell a single story: the cyber-offense cost curve just broke. Not bent — broke. And the policy apparatus is nowhere near calibrated for what happens when exploits become abundant rather than scarce.

Here’s what that means in practice. Critical infrastructure vendors who currently rely on “obscurity plus complexity as de facto defense” — meaning most of them — are now exposed in a way they weren’t eighteen months ago. The software liability question, which Washington has been debating in slow motion for years, just became urgent. And the insurance market for cyber events is going to either reprice dramatically or quietly start excluding AI-assisted attacks the way flood insurance started excluding certain coastal geographies. Watch that tell.

The exploit market pricing half a million dollars for a WordPress RCE was the canary. The canary just sang, and it cost $25 to teach it the tune. Scarcity was the entire security model. Build something else fast.