AI Developed a Zero-Click WeChat Worm in Nine Days. The Security Timeline Has Changed.

Zero-click exploits — vulnerabilities that require no user action, no link click, no incoming call answered — have historically been scarce and expensive to develop. They’ve been the domain of nation-state actors and well-resourced security firms because the research and weaponization cycle takes months of skilled human work.

Calif Research completed that cycle in nine days.

What Happened

WeWorm is a zero-click worm that exploits WeChat voice calls on iOS and Android. Infection occurs without the target answering or interacting with the device in any way. Calif Research used AI to identify the underlying bug and develop a remote code execution exploit in approximately two days. Worm development — writing the self-propagation logic — took an additional week.

Simon Willison highlighted the research and quoted the team’s own framing: “AI can already do most of the work here.”

The research itself is standard offensive security work — vulnerability discovery and exploit development are routine activities for legitimate security researchers, and the WeWorm disclosure follows responsible reporting norms. What is not routine is the elapsed time.

This is the second significant data point this year on AI-accelerated exploit development. In June, Anthropic’s own research showed Claude Mythos Preview building Windows exploit chains in under twelve hours. WeWorm adds a different dimension: a third-party research team working in a real-world mobile attack surface, building a complete propagating worm, in under two weeks.

Why It Matters

Security research has always operated on asymmetric timelines. Defending requires systematic coverage of the full attack surface; attacking requires finding one path through it. AI assistance compresses the attacker’s timeline without providing equivalent lift to defenders, who must still understand and remediate every vulnerability that gets discovered.

The more significant implication is team scale. Developing a zero-click exploit previously required a larger, specialized team with months of runway. Calif Research’s timeline suggests a small team with AI assistance can now operate in a window that serious security organizations previously couldn’t. The capability barrier hasn’t disappeared — this still requires security expertise and working knowledge of the attack surface — but the threshold has dropped.

What This Means for Builders

Capability evaluations are running behind reality. Published dangerous capability evaluations from frontier AI labs are accurate at the time they’re released. The WeWorm timeline is evidence that the combination of frontier models and skilled researchers can move faster than the evaluation publication cadence. Treat published evaluations as floors, not ceilings.

The “AI as accelerant, not capability” framing is outdated. Security frameworks that position AI as a tool for faster human work — rather than a contributor to what a small team can accomplish at all — are building on a premise the data no longer supports.

Defensive tooling deserves the same urgency. The same AI capability that accelerated WeWorm development can scan codebases for vulnerability patterns, generate security test cases against production endpoints, and surface attack surface that manual review misses. The question for platform teams is whether they’re applying it with the same urgency as the research is receiving on the offensive side.

Source: Simon Willison's Weblog

Content created with AI assistance and reviewed for accuracy.