Vulnerability disclosures have doubled over the past year to more than 10,000 per month, and AI is the engine behind it. That's the headline finding from Google's latest threat research, published Wednesday, which concludes that the pace and character of vulnerability discovery is fundamentally shifting — and not in a comfortable direction for defenders.

The number that should give you pause isn't the volume. It's the type. Google found that AI-discovered vulnerabilities are disproportionately likely to enable remote code execution — the top tier of severity, where an attacker can run arbitrary commands on your system without being anywhere near it. Human researchers and automated scanners have always found the low-hanging fruit first. AI is apparently reaching higher up the tree.

The mechanism makes intuitive sense. AI systems can fuzz code, trace execution paths, and reason about protocol interactions at a scale and speed that no human team can match. They don't get bored after the first thousand test cases. What used to require a specialist with deep knowledge of, say, DTLS protocol handling can now be approximated — imperfectly, but meaningfully — by a model running overnight.

Today's headlines are a live demonstration of what that looks like in practice. Citrix NetScaler is being actively exploited via a memory overflow in its DTLS implementation (CVE-2026-88772, CVSS 9.5). OpenSSL just patched a DTLS flaw that can leak heap memory. Chrome and Firefox shipped updates patching more than 100 vulnerabilities combined. Apple disclosed a zero-day already being weaponised in targeted attacks. That's four separate, high-severity disclosures in a single news cycle — exactly the kind of volume surge Google is describing.

The defender's maths is getting worse. Patching has always been a race, but it was a race with manageable odds. At 5,000 disclosures a month, a well-staffed security team could triage, prioritise, and action the ones that mattered. At 10,000 — with a higher proportion of RCE-class flaws — the same team is running faster just to stay in place.

What Google didn't say is equally significant. The same AI capabilities that find bugs can be turned toward exploiting them. The gap between disclosure and weaponisation has been shrinking for years; AI is likely to compress it further still.

For Australian organisations, the practical upshot sits squarely inside the ASD's Essential Eight: application patching at Maturity Level 2 requires patching "within one month," but at current disclosure rates, a month is an eternity for a critical RCE. The ACSC has previously recommended moving toward continuous patching cadences for internet-facing systems — that advice has never been more relevant.

What to watch: whether Google or other major vendors start publishing AI-assisted CVE counts as a distinct category. If defenders can see which bugs were AI-found, they can weight them accordingly in triage. Right now, that signal is buried in aggregate numbers — and the aggregate numbers are moving fast.