TL;DR
The National Cyber Security Centre’s operations director, Paul Chichester, told New Scientist that the four states the UK counts as its principal cyber adversaries, China, Russia, North Korea and Iran, now use AI to increase the output of their attacks, and that the NCSC uses commercial frontier models and its own small bespoke models to defend. The interview marks the centre’s tenth anniversary and follows a week in which MI5 announced an AI threat centre and the AI minister said a breach of UK defences “seems likely”.
Chichester’s description of the change is about tempo rather than novelty. “They don’t need sleep, coffee, food or anything like that,” he said of AI-assisted attackers, adding that their productivity “is certainly increasing”. Attacks, he said, feel “more relentless now than it did 10 years ago”, and cyber operations have become “an absolutely critical part of covert statecraft”.
Where AI enters the attack
According to Chichester, adversaries apply AI across what the NCSC calls the cyber kill chain: reconnaissance, developing a vulnerability and executing the intrusion. He also pointed to a ceiling that has gone. When people ran attacks, the volume of data or intelligence they could carry off had a practical cap; an AI-run operation has none. The motives vary by state, from espionage and economic advantage to disruption, with North Korea notable for outright financial crime and fraud.
New Scientist places the remarks against recent disclosures by large AI companies of models escaping test environments and attacking real systems, and the now-routine use of AI by penetration-testing firms, which can design new attack types and run them without pause.
The defence
Chichester would not say how far the UK conducts comparable operations, beyond noting “there are capabilities that are avowed and the UK uses”; New Scientist points to the National Cyber Force, run jointly by GCHQ and the Ministry of Defence. On defence he was more forthcoming. The NCSC makes no secret of using commercial frontier models, and it also builds small, task-specific ones. Training a frontier-scale LLM “takes a ridiculous amount of computing and machinery”, he said, but “building smaller models that do particular specific things is a very different proposition”, citing malware research and engineering investigations where AI has been in the tooling “for a long time”.
Looking forward
The interview adds an operational voice to a week of strategic warnings. We reported MI5’s director general setting up an AI threat assessment centre and AI minister Kanishka Narayan’s view that a Medicare-style breach of UK defences is likely. In our view the detail UK businesses should take from Chichester is the removed ceiling: an attacker that can exfiltrate without a human bottleneck changes the value of limiting what any one compromised account can reach, which is a question of architecture rather than of buying more detection.