Back to overview

    Two Weeks From Flaw to Worm — and No CVE Number

    A team at the security firm Calif used AI assistance to build a worm that takes over WeChat accounts through an unanswered call and propagates through the victim's contacts — cross-platform, with no interaction at all. The real story is the timeline: bug known on 23 July, working worm on 11 August. And there is no CVE number, no advisory, no catalogue entry.

    Published by J. Benjamin Espagné
    ai-securityzero-clickmobile-securityschwachstellenmanagementwechatlieferkettenis2exploit-entwicklung
    Share

    A contact you have not heard from in a while calls. You do not pick up. Seconds later your WeChat account has been taken over, the attacker is reading your messages — and is already calling your contacts.

    That is how Dustin Volz described the attack in the New York Times of 8 September. A small team at the Palo Alto security firm Calif built it. The worm is called WeWorm, exploits a memory-corruption flaw in WeChat's VoIP stack, and spreads with no user interaction at all — from Android to iOS and back, through each victim's saved contacts.

    WeChat and Weixin together reported 1.439 billion monthly active users as of 30 June 2026. Vinh Nguyen, formerly chief data scientist at the NSA, told the NYT the worm could have reached hundreds of millions of devices within hours.

    Tencent has closed the flaw. As far as anything public shows, it was never exploited in the wild.

    The number that matters

    Not 1.4 billion. Two weeks.

    DateEvent
    23 July 2026The team knows about the bug
    30 July 2026First working remote-code-execution exploit for Android
    11 August 2026Functioning worm demonstration, cross-platform
    21 August 2026Tencent ships 8.0.77 (Android) and 8.0.76 (iOS)
    28 August 2026Calif confirms the exploit is blocked server-side as well

    Calif puts the path from discovery to the first RCE exploit at around two days, and building the worm at another week. For an experienced team that used to be months of work.

    Calif's chief executive, Thai Duong, calls the bug exceptional and describes the combination of simplicity and power as the kind of thing attackers dream about.

    What has collapsed here is not the cost of an attack but the cost of developing one. That is a different quantity, and until now it was the most effective brake we had. Capabilities of this class belonged to state actors and a handful of commercial vendors — not because the knowledge was secret, but because the labour was expensive.

    The part the reprints leave out

    Three things get lost in the excitement over "AI builds worm".

    It was a defensive company. By its own account Calif builds such tools not to sell them but to improve defence. The finding went to Tencent through coordinated disclosure, the patch shipped before publication, and the demonstration followed afterwards. That is the sequence you want.

    The same capability found the flaw. The models did not merely build the exploit; they discovered the bug. Reading this case as proof that AI is an attacker's tool is reading half of it. The difference is not in the technology but in the fact that attackers do not practise coordinated disclosure.

    And several aggregators got the name wrong. A number of reprints refer to "California researchers". Calif is the company's name. Anyone reusing the story should check — it is an indication of how rarely anyone in that chain read the primary source.

    The real finding for defenders: there is nothing to prioritise

    This is where the case gets uncomfortable for vulnerability management.

    Tencent published no advisory. The release notes for the affected versions mention bug fixes. There is no CVE number. And the protection runs server-side, so no user has to install anything.

    From a defender's perspective that means this vulnerability exists in no catalogue. No NVD entry, no EPSS value, no KEV status, no line in any report. A process that tracks risk through CVE numbers — which is to say, essentially every process — never saw this one.

    We have described elsewhere how a flaw with a CVSS of 10.0 can land at zero in a prioritisation because asset context is missing. This case is the next stage: here it is not the context that is missing, it is the record.

    That is not an exception but the norm for mobile applications. Vendors patch quietly, app stores show "bug fixes", and distribution runs through auto-update rather than a patch cycle. It mostly works well — but it escapes the evidence trail NIS2 requires.

    What this means in practice

    WeChat is not a mass-market product in Europe. The case is relevant anyway, for three reasons.

    First, more organisations are directly affected than they think. Anyone doing business in China communicates over WeChat or WeCom — with suppliers, sales partners, local plants. Often on personal devices, often unmanaged. A hijacked account reads business correspondence and can write in its owner's name.

    Second, the pattern transfers. A memory-corruption flaw in the VoIP stack of a messaging app, triggered by an incoming call nobody has to answer, is not a WeChat peculiarity. The class has been known since the zero-click attacks delivered through WhatsApp calls, and it affects any application that processes incoming media before a human consents. Teams, Signal, WhatsApp, any softphone in the estate.

    Third, the application is part of the supply chain. NIS2 Article 21(2)(d) requires supply chain security. A communications app on a device carrying business traffic is part of that chain — even where nobody holds a contract with its vendor.

    What to do

    1. Build an application inventory for mobile devices. Which communication apps run on devices that carry business traffic, BYOD included? Without that list the rest is moot.
    2. Enforce updates rather than recommending them. Set minimum versions for critical applications through device management. Where that is impossible, the application is unsuitable for business use.
    3. Watch release notes, not only CVE feeds. For the handful of applications that genuinely matter, the version history is worth reading. "Bug fixes" after a report like this one is a signal.
    4. Restrict incoming calls where the application allows it. Zero-click over VoIP requires that the media stream be processed at all.
    5. Abandon the assumption that a patch cycle measured in weeks suffices. Two weeks from bug to cross-platform worm — and that was a team that reported voluntarily.

    What remains

    The reassuring reading: it was a defensive team, the flaw is closed, nothing happened. All true.

    The uncomfortable reading: precisely the same work is now within reach of anyone with two weeks and access to current models. And that person does not report.

    Between those two readings sits the question of whether your defence is built on warning time. If it is: that warning time was a by-product of high development costs on the other side. It just got cheaper.

    Frequently asked questions

    Do we even use WeChat in the business?

    More often than people assume. Organisations with suppliers, sales partners or production sites in China generally communicate over WeChat or WeCom — frequently on personal devices and outside device management. A mobile application inventory answers the question; an assumption does not.

    We do not use WeChat. Does the case still concern us?

    As a pattern, yes. The attack exploits a memory-corruption flaw in the VoIP stack, triggered by an incoming call nobody has to answer. The same attack surface exists in any application that processes incoming media before a human consents — including Teams, Signal and any softphone deployment.

    Why is there no CVE number for this flaw?

    Tencent published no advisory; the release notes mention only bug fixes. For mobile applications that is the rule rather than the exception. In practice it means a process tracking only CVE feeds never sees cases like this one.

    What does "blocked server-side" mean?

    Tencent additionally suppressed the attack on its own infrastructure, so no user has to install anything. That protects quickly and broadly — but it does not replace the app update, because server-side measures can be changed at any time and are not documented.

    What would be the single most important measure?

    Enforce minimum versions for communication apps through device management rather than relying on voluntary updates. Where that is organisationally impossible, the application is not suitable for business use.

    Sources