The rise of AI-powered cyber attacks: what you need to know

Every security vendor now has an opinion about AI-powered attacks, and most of those opinions are marketing. It is worth separating what has genuinely changed from what has simply been renamed.

What has actually changed

Phishing stopped being easy to spot

The traditional advice — watch for poor grammar, odd phrasing, generic greetings — is now close to useless. Language models produce fluent, contextually appropriate business correspondence in any language, at no meaningful cost. An attacker who previously sent one clumsy template to ten thousand people can now send ten thousand tailored messages, each referencing the recipient's actual role, employer, and recent public activity.

The consequence is not that phishing became possible. It is that the economics inverted. Targeted phishing used to be reserved for high-value targets because it took human effort. It no longer does.

Reconnaissance compressed

Mapping an organisation's staff, suppliers, technology stack, and internal vocabulary used to take an attacker days of manual work. That research is now largely automated. By the time a message arrives, the attacker may know your reporting line, your finance system, and the phrasing your company uses internally.

Voice and video became viable

Voice cloning is now convincing with a few seconds of source audio, which is trivially available for anyone who has spoken on a podcast, a webinar, or a recorded all-hands. Several large fraud cases in the last two years have hinged on a cloned voice authorising a transfer, or a deepfaked video call adding credibility to a fraudulent request.

Vulnerability research accelerated

Models are demonstrably useful at reading unfamiliar code and identifying suspicious patterns. This cuts both ways — it helps defenders too — but it shortens the window between a patch being published and a working exploit being available.

What has not changed

Almost everything else. The overwhelming majority of incidents we respond to still begin with one of the same few causes:

• Credentials without multi-factor authentication
• An internet-facing service missing a patch published months ago
• An over-permissioned account doing something it never needed the permission for
• A forgotten environment nobody was monitoring

None of these are AI problems. AI made the initial approach more convincing; it did not change what happens after the attacker is inside. The lateral movement, the privilege escalation, the exfiltration — these still look much as they did five years ago, and they are still caught by the same controls.

What to do about it

Stop relying on human detection of phishing. Training that teaches people to spot bad grammar is now training people to trust well-written fraud. Shift the emphasis to process: out-of-band verification for payment changes, and a culture where checking is never treated as an insult.

Deploy phishing-resistant authentication. Hardware keys and passkeys defeat credential phishing regardless of how convincing the message was, because there is no code for the user to hand over. This is the single highest-value control available.

Establish verification protocols for consequential requests. Any instruction to move money or change payment details should be verified through a channel the requester did not choose. A cloned voice cannot pass a callback to a number already on file.

Reduce your discoverable surface. The reconnaissance advantage only matters if there is something to find. Continuous attack surface monitoring closes the gap between what you think you expose and what you actually expose.

Assume compromise and invest in detection. Since the initial approach is harder to stop, the value of catching the second stage rises. Anomalous authentication, unusual data access, and unexpected outbound traffic remain highly detectable.

The honest summary

AI has made attackers faster, cheaper, and more convincing at the point of first contact. It has not given them new fundamentals. Organisations with strong authentication, a known asset inventory, and functioning detection are handling this shift comfortably. Organisations that were relying on staff to notice a typo are not.