decrypted · 29 october 2026 · vulnerabilities and patching · uk policy and law · ai and llm security

123 Reg says an AI-driven attack read its customer data

This morning's bulletin covered a schools platform recovering from ransomware. Since then, a UK domain registrar has told customers it was breached in what it calls its first AI-driven cyberattack. 123 Reg, owned by GoDaddy, says an intruder read customer records on 18 September. The claims about how the attackers worked are the interesting part, and it pays to separate evidence from the company's own assessment.

What was taken

The Register reported on 28 October that notification emails reached customers this week. A company spokesperson listed names, phone numbers, countries, language and currency preferences, account creation dates, products held and marketing preferences, plus company VAT numbers. Passwords and payment details are not listed. The company says the attacker only read the data, but copies may have been made.

A customer notice reported by VINnews describes GoDaddy Websites + Marketing profile and contact details reached through an API, the way one program talks to another, also dated 18 September. The dates match, but neither source gives a customer count or says how the attacker got in.

The AI claim

123 Reg says it found attack "skills" files on Chinese infrastructure, including copies of skills Anthropic had open sourced, and that the attacker used ARTEX. CrowdStrike published on 7 October that ARTEX is a recently released, open-source agentic penetration testing tool developed in China, used against South Korean financial firms, with the attacker's Claude Code session histories left exposed. CrowdStrike has not attributed that activity to a named adversary and assesses, with moderate confidence, a Chinese-speaking, financially motivated actor.

The link between that campaign and 123 Reg rests on 123 Reg's suspicion and overlapping timing. The China nexus is its own reading of attacker addresses in China and Singapore. Treat both as claims.

Think of an agentic tool as a tireless junior tester. You hand it a goal, it tries every door, reads what comes back and tries the next. It is cheaper than a skilled human, and patience stops being a scarce resource.

The design lesson

The one mechanism reported is an API that handed back customer data. Secure by Design asks different questions of it. Does each caller see only the records it needs? Is there a rate limit? Does anything alert when a client reads methodically, record after record, at three in the morning?

A read-only intrusion is quiet: no outage, no ransom note. The defender's only sign is in the logs, as a steady, systematic pattern of requests that no human customer would make. Scoping and rate limits are cheap at design time and expensive to retrofit.

For UK organisations that hold domains here, the practical risk is phishing built from real account details. Turn on two-step verification and log in through the registrar's website rather than links in emails. Opinion, not advice.

Also this week

Vulnerabilities: a low-severity flaw used as a zero-day. CrowdStrike's 2026 Legal Threat Landscape Report, dated 28 October, says the China-nexus group Warp Panda exploited CVE-2026-41709 as a zero-day, a flaw attackers use before a fix exists, "as late as early 2026". SecurityWeek describes it as a low-severity VMware ESXi logging flaw: an attacker logged in as administrator can perform certain operations without them being logged. Its value is concealment. SecurityWeek reports it joined CISA's KEV list, the catalogue of flaws known to be exploited, on 28 October, with US agencies given two weeks to patch, and that fixed ESXi builds exist for the 8.0, 9.0 and 9.1 branches. The same Broadcom batch included critical vCenter flaws, so patch the whole advisory. The design point: a severity score measures what a flaw does, not what it hides.

UK policy: Ofcom says four age-check methods can work at 16. Ofcom published its rapid assessment on 29 October, ahead of the Government's detailed rules for the under-16 social media ban. It names photo identity matching, digital identity services, banking-based checks and facial age estimation as capable of being highly effective at 16 or over. It says facial estimation has a margin of error of a few years at 16, so services should use a "challenge age" approach with a second check, and that circumvention "cannot be entirely eliminated". Ofcom says the methods can comply with data protection law and that it is working closely with the Information Commission. The Government is due to set out how the ban works by the end of 2026, with the ban in force in 2027. The question that matters is who holds the proof of age afterwards. A design that confirms "over 16" and keeps nothing else is the one UK organisations should ask their vendors for.

Sources