How AI Is Transforming Cybersecurity — For Attackers and Defenders

AI hasn't created a new category of cybersecurity risk so much as it's changed the speed and economics of the categories that already existed. The techniques attackers use — phishing, social engineering, reconnaissance, credential abuse — are largely the same ones security teams have defended against for years. What's different is that AI has collapsed the cost and skill required to run them well, at the same time it's giving defenders new ways to keep up. Both sides of that shift matter, and EMEA organizations are dealing with a third layer on top of it: a regulatory framework — the EU AI Act — that's still actively changing shape.

What's actually different on the attacker side

The realistic threat isn't "AI discovers a zero-day no human could find." It's that generative AI removed the tells that used to make phishing easy to catch — bad grammar, awkward phrasing, an obviously translated tone — and made producing a convincing, personalized lure as fast as producing a generic one. It's also made voice and video impersonation viable at a cost that used to be prohibitive. The most consequential real-world example so far is Arup, the British engineering firm that lost HK$200 million (about US$25.6 million) in February 2024 when a finance employee in its Hong Kong office was convinced to make fifteen separate wire transfers after a video call with what appeared to be the company's UK-based CFO and several colleagues — all deepfakes, built from publicly available meeting footage. Arup's own CIO was direct about what this actually was: not a systems breach, but "technology-enhanced social engineering." No network was compromised. A person was.

What's actually different on the defender side

The same properties that make AI useful for generating a convincing lure make it useful for triaging the alert volume a modern SOC actually deals with — pattern recognition across a scale of log and telemetry data no analyst team can manually review, faster correlation across signals that used to require chaining several separate tools by hand. This isn't a replacement for a security team's own judgment, and it isn't a magic detector for "AI-written" content (that detection problem remains genuinely unsolved at any reliable accuracy). It's a force multiplier on the parts of the job that were always about volume and speed — which is exactly the part attackers are also now accelerating.

The regulatory layer EMEA organizations specifically have to deal with

The EU AI Act entered into force in August 2024, and its timeline has already changed once since then. The original date for high-risk AI system obligations under Annex III (the use-based category most enterprise AI deployments fall under) was August 2, 2026 — but a May 2026 political agreement, the Digital Omnibus on AI, pushed that deadline back to December 2, 2027. Product-regulated high-risk systems (Annex I — things like AI embedded in medical devices or lifts) moved from August 2027 to August 2028. What didn't move: conformity assessment, technical documentation, CE marking, and EU database registration were still expected to be substantively underway well before the original August 2026 date for anything already in scope, and the penalties for getting this wrong are real — up to €35 million or 7% of global annual turnover, the same order of magnitude as GDPR's own maximum fine. If your organization has deployed or is deploying an AI system that touches EU users, this is now a compliance question sitting alongside GDPR and NIS2, not a separate track.

What this actually means day to day

The practical takeaway isn't "be afraid of AI" in either direction — it's that the fundamentals still hold, they just need to be applied faster and more consistently than before. Verify unusual payment requests out-of-band, through a channel the request itself didn't suggest — a phone call to a known number, not a reply to the same email or video call thread. Treat any AI system your organization deploys as something with a real, testable attack surface, not just a compliance checkbox. And if you're evaluating whether an AI feature you've built or bought actually holds up under adversarial pressure, that's a distinct kind of assessment from a standard penetration test — see our piece on what actually happens when someone tries to break an LLM-powered feature for what that testing looks like. Our AI Services and Security Consulting offerings cover both sides of this: testing what an AI system can be made to do, and helping map what EU AI Act obligations actually apply to your specific deployment.