A simulated power plant, a virtual industrial control system, turbines, pressure systems and temperature controllers: all virtual, yet realistic enough to serve as a benchmark for measuring the most dangerous capability of modern AI: autonomous cyberattacks.
The test is called “The Cooling Tower”. The UK’s AI Security Institute (AISI) uses it to measure what AI models can accomplish without human guidance. No model had ever succeeded, until last week. Claude Mythos, Anthropic’s latest model, completed the attack in three out of ten attempts. A first.
The growing capability of autonomous AI attacks
AISI is not a marketing organisation but an independent government body that evaluates frontier AI models before their wide-scale release. Its methodology is simple: how long can an AI model autonomously carry out a complex attack without getting stuck?
In November 2025, that capability doubled every eight months. By February 2026, it was already doubling every 4.7 months. Since then, Mythos and OpenAI’s GPT-5.5 have already surpassed that trend. AISI writes cautiously: “It is unclear whether this represents a new trend or a one-off jump.” But also: “The direction of change has always been the same.”
An attack that would take a human security expert twenty hours can now be carried out autonomously by Mythos.
The impact on the SME landscape
The AISI simulations are not replicas of attacks against large corporations with full-time security teams. They simulate attacks on small enterprise networks. Exactly the kind of environment that reflects the reality of most SMEs in Flanders.
The threat model is changing fundamentally. In the past, a targeted attack required specialist hackers and an attractive target. Today, attacks are scalable and automatable. Any SME with an online presence, a cloud-based ERP system or integrations with suppliers and customers is a potential target, not because of who they are, but because of how they are configured.
At Palo Alto Networks, the latest AI models carried out as many security tests in three weeks as had previously been performed manually in an entire year. The same technology that defends can also attack.
Priorities for business leaders and IT managers
The Centre for Cybersecurity Belgium (CCB) and the NCSC are clear about the basic measures: regular updates, strict access controls, multi-factor authentication and active logging. This is the minimum line of defence, yet for many SMEs it is still not a reality.
There is also good news from the Flemish side. Since 1 February 2026, cybersecurity consultancy has been the only remaining area for which SMEs can apply for subsidies through the SME e-wallet scheme. Up to 30% support, amounting to a maximum of €7,500 per year. The Flemish government has sent a clear political signal. An SME owner who does not make use of it is leaving both money and protection on the table.
The cooling tower stood. Now it doesn’t. The question is not whether your network will ever be scanned. The question is what will be found when it is.
Cybersecurity as a strategic business risk
AI-driven cyber capabilities are growing faster than most organisations can keep up with. That is not a reason for panic, but a call for structural action. Security is not an IT project. It is a business risk that belongs at board level.
