AI cyber attacks: your questions answered

They are attacks that use AI to improve phishing and impersonation, to find and exploit software flaws faster, or that target AI systems themselves. The NCSC expects AI to make parts of intrusion operations more effective[1]. Prompt injection is an example of an attack on AI systems[2].

Mostly they improve existing methods. The NCSC expects AI to increase the volume and impact of intrusions by enhancing existing techniques rather than creating novel threat vectors[1]. The Five Eyes agencies nonetheless say AI is shrinking the time between discovery and exploitation[3].

An employee joined video calls with impersonators of senior company officers, who asked for money to be sent to specified accounts, and HK$200m (about £20m) was transferred[7]. Arup said its internal systems were not compromised[7].

Yes. NCSC guidance says threat actors are increasingly using AI to discover and exploit vulnerabilities[4]. The NCSC's 2027 assessment expects AI-assisted vulnerability research to be the most significant development[1].

It is an attack that gets an LLM to treat hostile text as instructions. OWASP ranks it first in its 2025 list of LLM application risks[8]. The NCSC says it may never be fully mitigated and should be managed by limiting likelihood and impact[2].

The 2025/2026 government survey found that 43% of UK businesses and 28% of charities had identified a breach or attack over the past year[5]. The survey does not attribute those attacks to AI.

Few do. Among organisations using or adopting AI, about a quarter of businesses (24%) and charities (27%) had practices to manage AI-related risks[5].

The NCSC says it cannot be fully mitigated with a product or appliance, and advises scepticism towards vendors claiming to stop it[2].

The Five Eyes agencies ask leaders to understand and assess risk, readiness and accountability, prioritise foundational controls and give cyber leaders authority and resources[3]. See the defence plan.