How to counter AI-driven cyberattacks?

Cyber attackers are now handing a growing share of their operations to AI, and some campaigns reach a high degree of automation.

To counter AI-driven cyberattacks, companies need to speed up their detection and response capabilities, relying on contextual behavioural analysis and automated, targeted and reversible protections.

In an espionage operation described by Anthropic, an AI agent reportedly carried out 80 to 90% of tactical operations against around thirty organisations. Human operators stepped in at key moments to authorise the exploitation of a flaw or choose the data to exfiltrate, but the AI carried out most of the attack.

The major difference between a human attack and an AI-led attack is the speed at which actions are chained, along with the ability to try several paths and start again after a failure. The techniques themselves remain familiar: reconnaissance, credential theft, privilege escalation, movement across the network.

Faced with this threat, how should organisations' protection evolve to prevent this type of attack?

Defensive AI: an answer to automated cyberattacks

An organisation may already have the right detection tools yet still be too slow to act. Qualification, analysis, validation, deployment: several hours can pass between the first signal and effective protection.

Seemingly harmless events need to be connected, a suspicious sequence recognised and a measure applied only to the device concerned.

Defensive AI can help with this work, provided its decisions are framed: knowing the context, making the action proportionate to the risk and allowing teams to understand what was done, undo it and take back control.

Locking everything down further does not solve the problem. Overly broad protection blocks legitimate use, multiplies exception requests and often ends up being relaxed.

EDRs remain essential on endpoints, but their visibility depends on how widely they are deployed and benefits from being complemented by network traffic analysis, in particular to cover agentless devices.

An example: Cato Networks Agentic Threat Prevention

Thanks to its SASE architecture, Cato Networks brings networking and security together on a single platform. Its agents therefore have the context needed to analyse behaviour and apply protections directly to traffic.

The Agentic Threat Prevention module uses this architecture to match the company's activity history with signals from the security engines. It spots suspicious behaviour and applies temporary protections designed to stop an attack from progressing.

It is activated from the management console, with no additional deployment on endpoints. Settings allow automatic blocking or monitoring to be chosen for each threat category.

Every four hours, the agents analyse traffic and security signals from the previous twenty-four hours. This analysis defines the protections, which are then applied to traffic in real time.

An example illustrates how it works. An endpoint downloads a script from a code-sharing site. This action can be perfectly legitimate. But, combined with other clues, it may suggest the preparation of malicious remote access. A targeted response then consists of preventing the download of the tool that would enable this access, on that specific endpoint.

Another case: an Internet-facing server starts exploring internal resources it usually has no interest in. Restricting its access to certain sensitive services can stop the attack from progressing, without cutting connections for an entire site.

These restrictions are reassessed and can be lifted once the elements that justified them have disappeared. This avoids keeping rules that have become useless over time.

Cato's unified management console shows the suspicious behaviour detected, the controls activated and the risky actions actually blocked, as well as the number of devices concerned. For each device, a timeline and summary explain the detections, the protections applied and how they evolved, until they were lifted.

Deploying and operating defensive AI in the enterprise

The quality of decisions depends on what the agent sees. A common platform for networking and security makes it easier to analyse traffic and enforce controls on it. But incomplete inspection or poorly managed exclusions reduce this capability.

At SASETY, we help companies prepare and operate these defences. In particular, we check that encrypted traffic inspection is enabled and that its exclusions are relevant. We then define with the client, for each threat category offered, the mode suited to their business constraints: automatic blocking or monitoring.

The agent's actions are reviewed during security reviews. Where necessary, specific cases are analysed to adapt protections to business constraints.

Making defensive AI an operational reality

Faced with attacks that can adapt while they run, protections must be able to evolve too. Defensive AI opens up this path, provided it relies on a good knowledge of the environment and on decisions that teams can understand and control.

For companies, the challenge now is to bring these capabilities into their day-to-day security, with the same reliability requirements as their other means of protection.

SASETY supports companies with the architecture, deployment and operation of their network and cybersecurity infrastructure: more than 80 clients in 35 countries and more than 70,000 protected users.