AI social engineering: your questions answered

It is social engineering, persuading people to hand over money, access or data, with AI used to write messages, imitate people or support the attack. The NCSC notes that attackers already use AI to help gain entry to systems this way[1].

Phishing is the most common attack: 38% of UK businesses experienced it in the 2025/2026 government survey[4]. The figure shows scale, not AI's share, because it does not say how many involved AI.

Do not rely on spelling or grammar. The NCSC recommends checking anything important by phone or by logging in to the account directly[3]. Judge the request, not the writing.

Often not. In Vodafone's research 40% of respondents could recognise voice-based scams, against 63% for text phishing[8]. Use call-backs on numbers you already hold for risky requests.

No. The NCSC advises layered defences, including multi-factor authentication and incident response, and not relying on training alone[3]. In Vodafone's test, 33% of employees identified fraudulent communications[8].

It is an attack where crafted input makes an AI model act unexpectedly, such as leaking confidential information[9]. The NCSC doubts it can be removed entirely and advises cutting its likelihood and impact[2].

Anthropic reported a data extortion campaign against at least 17 organisations in which the actor threatened to expose stolen data and sometimes demanded more than $500,000[5]. That is a single reported case and says little about frequency.

It sees AI speeding up and sharpening some intrusion tasks, by refining familiar methods[1]. It also foresees a widening gap between organisations that keep pace and those that do not[1].

Make an inventory, limit what each tool can access, log its activity and ask suppliers about their security[2]. A government code of practice for securing AI sets baseline principles[6].

Few do so formally. The 2025/2026 government survey found that 24% of the businesses (31% of the total) already using or weighing up AI had practices for managing its risks[4].