Two Days to a Worm

Somebody just built a self-spreading, zero-click worm for an app with 1.4 billion users. It took them about two days. AI did the heavy lifting.

That's the part that should stick. Not the CVE. Not the vendor name. The timeline.

What happened

This week a research outfit called Calif went public with WeWorm, a zero-click worm for WeChat. The Register has the write-up. The bug is memory corruption in WeChat's VoIP stack. Someone on your friends list calls you. You don't have to pick up. If the call rings long enough, or you answer, the exploit fires. Seconds later they own your account: they can read your messages, send as you, and call as you. Then the compromised account starts calling the next person on the list. Decline the call and you're safe. Step away from your phone and they just try again.

Works on iOS and Android. Tencent confirmed it and shipped patches on August 21. Calif is holding the technical details for a conference talk, but they released a demo, and they were clear about why they went public early: to show this kind of capability is no longer limited to well-funded, sophisticated actors.

The line I keep re-reading: Calif said it used AI to find the vulnerability and develop its first remote code execution exploit in about two days.

Two days. From “let's look at the VoIP code” to a working cross-platform RCE that spreads itself.

We've been here before, slowly

Worms aren't new. The Morris worm knocked over roughly a tenth of the internet in 1988. SQL Slammer in 2003 infected most of its vulnerable hosts in about ten minutes and took out ATMs and airline systems, with no payload at all. Just traffic. WannaCry in 2017 shut down hospitals across the UK. NotPetya the same year cost Maersk and Merck hundreds of millions each, and it wasn't even aimed at them.

Every one of those took real skill and real time to build. That was the friction that kept the count low. A handful of teams on earth could write a reliable zero-click, self-propagating exploit, and most of them worked for governments.

That friction is gone. Or going. Pick your tense.

The math changed, not the malware

This isn't about one messaging app. In April, Anthropic pointed its Claude Mythos model at open source software under a program called Project Glasswing. Within two months it had flagged more than 10,000 high or critical vulnerabilities across a thousand-plus projects, and over 90 percent of a sampled batch checked out as real. One of them was in wolfSSL, a crypto library sitting in billions of devices.

Finding bugs used to be the expensive part. It's now the cheap part. The expensive part is fixing them, and that side of the equation still runs on volunteer maintainers, 90-day disclosure windows, and enterprises that patch quarterly if they patch at all.

So picture the gap. On one side, a model that turns a memory bug into a worm in 48 hours. On the other, an update cycle measured in months and a user base that never installs it anyway. That gap is where the next big incident lives.

Why I think this reaches the internet itself

I'll say the quiet part. I think AI-built worms can take down the internet as most people experience it, along with the critical communications riding on top of it.

Not because the AI is evil. Because the targets are monocultures. One messaging app with 1.4 billion users. One VoIP stack. One TLS library in billions of devices. One BGP implementation, one DNS resolver, one SCADA protocol handler. A worm doesn't need to be clever to be catastrophic. It needs a bug in something everybody runs and a way to spread without anyone clicking. WeWorm had both, and finding it was cheap.

Now stop thinking about messaging apps. Point the same two-day process at a carrier's core network gear, a satellite terminal, the firmware in a common cellular modem, or the dispatch software every 911 centre in a province happens to run. Comms don't fail gracefully. When the messaging app is down you use the phone. When the phone is down you use the radio. When the radio dispatch software is down you're back to 1975 and nobody trained for it.

Slammer did its damage with 376 bytes and no payload, in ten minutes, in 2003. Give the next one a model that can find a fresh bug in whatever it lands on.

The caveat

This doesn't happen next month. AI finds bugs faster than it writes stable, weaponized, self-propagating exploits, and Calif were skilled humans steering the tool, not a prompt left running overnight. The number of people who can take an AI-found crash and turn it into WeWorm is still small. Smaller than last year, but small.

And defence gets the same tools. Glasswing found 10,000 bugs so they could be fixed. wolfSSL is patched. Tencent closed WeWorm before it was public.

But look at the asymmetry. The attacker needs one bug in one popular thing. The defender needs every maintainer of every popular thing to patch inside the window, and every user to actually apply it. One side just got a speed boost the other side can't match by working harder.

What to do about it

If you run an organization, the advice is boring and I'm giving it anyway, because the boring stuff is what's going to matter.

Assume the patch window is days, not quarters. If your change process can't push an emergency patch to phones, endpoints and network gear inside 72 hours, fix the process before the worm shows up, not after.

Kill the monoculture where you can afford to. Two carriers. Two messaging channels. An out-of-band way to reach your incident team that doesn't depend on your identity provider or your phone system. In 600-plus incident responses I've watched more companies lose hours because the comms tool was the thing that went down than I care to count.

Practise being offline. Tabletop the scenario where the internet is degraded for three days, cell service is flaky, and your MSP is dealing with forty other clients. Who can make decisions? How do you reach staff? Where's the paper copy of the call tree?

Push your vendors. Ask what their patch SLA is for a zero-click RCE. Ask how they're running AI against their own code before someone else does. If they can't answer, that's your answer.

And for the industry: stop treating “the AI found a bug” as a press release. Every one of those findings is a clock that started ticking.

Bottom line

A worm for a 1.4 billion user app got built in two days with AI in the loop, and the vendor had it patched before anyone outside the lab knew. That's the good version of this story. The bad version is the same two days pointed at something that can't be patched over the air, by someone who doesn't file a disclosure.

The tools that find the holes get faster every quarter. The people who fill them don't.