decrypted · 14 november 2026 · ransomware and cybercrime · supply chain · ai and llm security

A record quarter of cyber incidents, and the regulator will chase almost none

Britain's data regulator has logged more cyber incidents in one quarter than ever before, and it will investigate almost none of them. The Information Commissioner's Office (ICO), the UK's data regulator, received 1,371 cyber incident reports between July and September, according to its figures as analysed by The Record. That beats the previous three-month high, set in 2023.

What the numbers say

The figures are provisional, and they only count incidents that organisations discover and choose to report. Across all data security incidents, including dull ones like an email sent to the wrong person, the quarter hit a high of 4,604.

The sector movement is telling. Retail and manufacturing rose from 205 to 296 reports between the second and third quarters. Charities nearly doubled, from 79 to 145. Education was the one big sector to fall.

Why ransomware looks quiet

Here is the odd bit. Ransomware reports to the ICO are falling: 1,253 in 2023, 752 in 2024, 620 in 2025, and 321 in the first nine months of 2026. Meanwhile Sophos counted criminal groups naming more than 1,200 new victims on leak sites worldwide in October, its highest monthly total, and about 350 British organisations have been named this year, roughly 50% more than last year.

Both can be true. Think of the ICO's ransomware box as a smoke alarm tuned to one kind of fire. Many modern gangs lock nothing. They steal the data and threaten to publish it, so no screen goes dark and nobody files the incident as ransomware. The Record notes it is unclear whether the ICO's category captures that extortion, and most new attacks are filed under "other cyber incidents", a catch-all.

What the regulator does with it

The ICO pursued investigations in just 18 of the 1,371 cyber incidents, and most cases closed with informal action. No regulator can chase every report, so your resilience is yours to design, not something enforcement will provide.

Jonathon Ellison, the director for national resilience at the NCSC, the National Cyber Security Centre, said many successful attacks still exploit known and preventable weaknesses, with AI increasing the speed and scale at which they can be found.

The Secure by Design lesson

If the attack is theft rather than encryption, the design question changes from "can we restore?" to "what could be taken?". Backups answer the first and do nothing for the second.

Hold less. Data you deleted last year cannot be published this year. Retention limits cost a policy and an afternoon.
Separate it. One stolen login should not reach every record, so split data stores by purpose and by who needs them.
Close the known doors. Sophos recommends phishing-resistant multi-factor authentication, a second check on top of the password, plus passkeys, and fixing flaws in internet-facing systems.

The test for a UK director: if a thief walked out with your busiest system tonight, could you say by Monday exactly what they took?

Also this week

A tampered wallet, from a reseller. Trezor says three Safe 5 hardware wallets bought from the reseller CryptoBilis between 20 August and 23 September 2026 contained an extra circuit board that copied the screen, including recovery words, and sent it over the mobile network to the attacker. Trezor says its own firmware and Secure Element were not altered, and that 48 CryptoBilis customers have contacted it, with no reports of lost funds. It released firmware 2.13.1 on 11 November, which stops the implants it found capturing anything, but it cannot protect recovery words captured earlier, so affected owners should move funds to a new wallet with a new seed. Trezor has also cut its reseller network by about 80%. The lesson for buyers of any security kit is provenance: buy through the maker's own channel, and treat the supply chain as part of the product. Hardware you cannot inspect is trust you cannot verify.

A prompt injection that reached a publishing pipeline. Researchers at Depi described chaining a hostile GitHub issue, which an AI agent in OpenAI's Codex repository read as instructions, into stolen build tokens and a poisoned cache, with a route towards publishing a backdoored npm package, JavaScript's public library of add-on packages. The workflow let anyone who could open an issue trigger the agent. Its sandbox had no network, but leftover debug pipes exposed the memory of the build runner beside it. It was a proof of concept on Depi's own test environment, not reported in the wild, and OpenAI fixed it and paid an initial $1,000 bounty. The design flaw was letting an AI agent read untrusted text while holding powers anyone could trigger. If your pipelines let a chatbot read public input, give it nothing worth stealing, and scope tokens so that one workflow's cache cannot feed another with more privilege.

Sources

If you want help working out what a thief could take from your systems, get in touch.