decrypted · 17 october 2026 · ransomware and cybercrime · vulnerabilities and patching · ai and llm security

ASOS shows how one convincing phone call becomes a customer data breach

ASOS has now said how its breach began, and the answer owes less to clever malware than to a convincing stranger. The retailer told customers on 8 October that an attacker posed as a trusted contact to obtain an employee's login details, then used them on third-party platforms ASOS relies on to talk to its customers. Ten days on, it is a tidy case study of a pattern every UK board should recognise: the weak point was a person, but the weakness was a design choice.

What ASOS has said

On 6 October a push notification reached customers through ASOS's own app, claiming the company had been hacked and alleging that a Snowflake environment was compromised. ASOS said no customer passwords or payment card information were accessed. Snowflake said the incident did not result from a vulnerability, weakness, flaw or misconfiguration in its service. The BBC reported that a group calling itself Xuanyewen claimed to have used Simon AI, a marketing platform built on Snowflake. ASOS has not confirmed that, so treat it as a claim. The BBC also reported that a sample of the stolen data went beyond the "basic contact details" ASOS first mentioned, including addresses, customer numbers and search terms. ASOS has not said how many customers are affected.

The mechanism, in plain English

Think of a hotel where one master key opens every floor, and the front desk hands it over to anyone who rings up and sounds like head office. The employee's login was the key. The marketing platform was the floor holding the customer file. Nothing here needed an exploit or malware, which is why it is hard to catch. A defender would see a valid sign-in, then a large pull of customer records, then a message sent by a legitimate tool. Spotting it means watching for unusual volume and unusual sending, not for malicious code.

The Secure by Design lesson

Four design decisions would have changed the outcome, and none of them is exotic:

Phishing-resistant sign-in. Passkeys or hardware security keys for anyone who can reach customer data. A persuasive caller can talk someone out of a password. They cannot talk a hardware key out of a drawer.
Least privilege. A marketing login should not be able to export complete customer profiles, and a bulk download should need a second approval.
Two people to message everyone. Sending to the whole customer base through your own app should never be one account's decision.
Suppliers' platforms are your estate. Apply the same sign-in rules and keep the logs where you can read them.

The cost is mostly friction and a rollout project, not new technology. Set that against customers hearing about your breach from your own app.

Also this week

Drupal fixes a critical remote code execution flaw. Drupal published fixes on 14 October for CVE-2026-108097 in core's Field UI module. Remote code execution means running your own code on someone else's machine. The project rates it Critical, 19 out of 25, and says core alone is not enough: contributed modules or custom code provide the routes in, and at least one needs no login. Fixed releases are 10.6.19, 11.3.19 and 11.4.10, and older branches get no fix. A related Entity List module flaw, CVE-2026-108099, is fixed in 3.0.10. Drupal rates exploit maturity as "Theoretical" but warns exploits might appear within hours or days, and BleepingComputer reports no confirmed exploitation. If you cannot update yet, uninstalling Field UI is the stopgap. Sites still on Drupal 8 or 9 are outside support altogether, which is design debt, not a patching delay. If your site, or your agency's, runs Drupal, ask which branch it is on.

Anthropic warns small AI models may soon spread across cheap devices. Anthropic's Frontier Red Team published research on 12 October saying small models may be "months away" from copying themselves onto connected everyday devices such as off-brand Android TV boxes. In the lab, a 9B model got inference running on eight of nine hardware types, though it was slow, taking more than a day on the hardware tested. The test was simplified, using an unauthenticated debugging port, and the authors call the results an upper bound. No self-spreading AI has been shown on real hardware. Their defences include internet provider monitoring and secure device defaults, and they cite the UK's ban on default passwords on connected devices as the kind of regulation that shrinks the pool of easy targets. That is Secure by Design doing quiet work.

Sources