The number is almost too large to process: 271 previously unknown, exploitable security vulnerabilities — all found in Firefox, all discovered by a single AI model, in a research collaboration between Mozilla and Anthropic. Bruce Schneier flagged it this week, and it's the kind of number that makes experienced security researchers do a double-take.

For context: a skilled human researcher finding a handful of zero-days in a mature, well-audited codebase like Firefox in a year would be considered exceptional work. Mozilla's earlier AI collaboration with Anthropic — using Claude Opus 4.6 — had already turned up 22 security-sensitive bugs, enough to prompt fixes shipped in Firefox 148. Then came Claude Mythos Preview, an early access model, and the numbers exploded.

What's actually happening here is that AI systems are getting good enough at code comprehension and reasoning to do the hard, tedious, expert work of vulnerability research at scale. This isn't AI writing better phishing emails. It's AI doing what previously required a team of elite reverse engineers working for months — and doing it faster, cheaper, and apparently more thoroughly.

The implications split in two directions, and neither is comfortable.

On the defensive side, this is legitimately great news. Mozilla found these bugs and can fix them before attackers do. That's the system working. Security teams at well-resourced organisations can now deploy AI-assisted fuzzing and static analysis to find their own flaws before adversaries do. Wiz used a similar AI-assisted reverse-engineering approach to uncover the critical GitHub RCE vulnerability (CVE-2026-3854) disclosed this week — a bug previously considered too costly to hunt manually.

On the offensive side, the same capability is available to anyone with access to a frontier model. If Anthropic's model can find 271 bugs in Firefox, what's stopping a well-funded threat actor from pointing equivalent tooling at critical infrastructure software, medical systems, or industrial control platforms? The answer, for now, is mostly cost and access — thin moats that are eroding quickly.

The LiteLLM story from this week is instructive as a near-term preview of that world. CVE-2026-42208, a critical SQL injection flaw in the popular open-source LLM gateway, was disclosed — and actively exploited — within 36 hours. Attackers are already moving at AI speed. Defenders need to as well.

For Australian organisations, the Firefox finding lands in a context where the ASD's Essential Eight and ACSC guidance both emphasise patching as a primary control. The Essential Eight's Patch Applications control exists precisely because browser vulnerabilities are a top initial-access vector. When a single AI run can surface 271 of them in one codebase, the patch cadence conversation changes — because the volume of disclosed CVEs is about to get much, much louder.

What to watch: Whether other browser vendors and major open-source projects announce similar AI-assisted audit programmes. If Mozilla is getting 271 bugs per AI run, the queue of critical patches heading toward users in 2026 could be unlike anything we've seen. That's ultimately good — but the transition period, where bugs are found faster than they can be fixed and communicated, is going to be messy.