AI powered cyber attacks: your questions answered

It is an attack where criminals or state-linked actors use AI to scale or improve their methods, such as phishing, impersonation and intrusion. The NCSC expects AI to keep making parts of an intrusion quicker and more effective[1]. Attacks on AI tools, such as prompt injection, are a related but separate category[2].

The 2025/2026 government survey counted 43% of businesses and 28% of charities as having had a breach or attack in the past year[6]. The survey does not attribute those figures to AI[6].

An employee was tricked into transferring HK$200m (about £20m) after a video call in which synthetic voices and images impersonated senior officers[8]. The company said its own systems had not been breached[8].

Anthropic reported a campaign it assessed as Chinese state-sponsored in which AI did 80-90% of the work, with humans making perhaps four to six critical decisions per campaign[5]. The company describes it as the first documented case of a large-scale attack executed without substantial human intervention[5].

It is an attack that uses crafted text to make an LLM behave in ways its owner did not intend. OWASP ranks it first in its 2025 risks for LLM applications[9]. The NCSC says it probably cannot be fully prevented, so the aim is to limit likelihood and impact[2].

The NCSC advises being sceptical of any product that claims it can stop prompt injection[2]. It recommends designing the system so a fooled LLM has little to abuse, for example by limiting its privileges[2].

Not widely, according to the 2025/2026 survey. Among businesses that use AI or are moving towards it, 24% had practices to manage its risks, against 27% of charities[6].

It says security should be a core requirement across an AI system's lifecycle and that leaders should understand the consequences if its integrity, availability or confidentiality is compromised[4]. A separate government code adds baseline principles for those who develop and deploy AI[10].

The sources do not measure that directly. The NCSC judges AI will make elements of intrusion more effective[1], and phishing already led the UK survey, reaching 38% of businesses[6]. Treat message quality as an unreliable signal and rely on second-channel checks.