AI-driven attacks hit broadly, and confidence in autonomous tools falls

A CDW report states that 43% of organisations have encountered phishing enhanced with artificial intelligence, and that 37% have encountered malware driven by the same technology. The report also states that 41% plan to adopt AI for threat detection and anomaly detection. ZDNet covered the figures on 27 July 2026. A separate report from Cobalt states that only 9% of organisations are willing to rely on autonomous security tools in 2026, down from 29% in 2025. Dark Reading covered that report on 27 July 2026, and links the decline partly to adversaries exploiting governance rules rather than technical vulnerabilities.

What happens technically

AI-enhanced phishing is not a new attack path. It is a cheaper and better version of an old one. Language models write flawless English at scale, and they can tailor each message to the recipient's role, employer and recent activity. That removes precisely the signals many awareness programmes taught people to look for, namely language errors, clumsy phrasing and obviously wrong addressing. Buck Bell, director of CDW's Global Security Strategy Office, says we should be concerned about the increasing ability of AI as a technology to discover and exploit weaknesses. His advice in the coverage is that security practitioners must now use similar tools to find those weaknesses before attackers do.

The second report points to an entirely different mechanism. Autonomous security tools act within limits the organisation sets itself, such as how much traffic they may block, which accounts they may disable, and how far they may go without a human approving. Such caps are sensible practice, because a tool that can halt production systems needs a boundary. Independent security researcher Burak Oktenli calls it "cap weaponization" when adversaries exploit the governance rules themselves instead of a technical vulnerability. The coverage does not detail how this is done in practice, and it does not state which tools were examined.

What this means for you as the one responsible for security

Read the figures as a direction, not as a precise measurement. Neither sample size, geography nor question wording appears in the coverage. Both reports also come from vendors in the same market, which is a standard caveat for vendor surveys. We nonetheless recognise that direction in Norwegian organisations. The messages your people receive are better written, better targeted and harder to tell apart from genuine correspondence.

The consequence for your training is direct. If your programme teaches staff to look for language errors, it trains them on a signal that is disappearing. What still holds is verifying the request itself in a channel other than the one it arrived in. The fall in confidence in autonomous tools points to something you should note as well. Automation, as we see it, is not the problem in itself. The limits it operates within have instead become part of the attack surface, and they deserve the same protection as the rest of your security configuration. The NIS2 directive places responsibility for such assessments with management in essential and important entities.

Berigo recommends

• Replace training that teaches staff to hunt for language errors with routines that verify the request in a channel other than the one it arrived in.
• Require confirmation outside the channel for changes to payment details, access rights and supplier contacts.
• Write down what limits your automated security measures have, who may override them, and who may read that document. Treat it as sensitive.
• Rehearse the case where the autonomous tool is wrong or fails to act, and how people take over.
• Put AI in both attack and defence into the risk register, and report it to the board alongside the rest of the risk picture.