AI agents are already capable of breaking into systems themselves: cyber experts warn

A few days later, Anthropic also issued a similar announcement, when an internal review of 141,006 test runs revealed that Claude models hacked, without explicit guidance, into the systems of three different organizations – and none of them noticed the hack themselves.

These are no longer theoretical laboratory exercises. This is living proof that artificial intelligence agents are already able to act independently today, find loopholes and exploit them – without human contact and without anyone noticing. The question that arises from this is who will protect us when the attacker is a machine that learns in real time. The answer, according to executives in the cyber industry, is – another machine.

An employee who can perform malicious actions

The significance in the cyber world is dramatic: operations that previously required skilled staff and hours or days of work can be performed by AI in minutes and on a massive scale. The time between discovering a weakness and trying to exploit it is getting closer to zero.

The first step for organizations is to understand at all what AI entities are operating within them. Which agents, models and tools are connected to the organization’s systems, what information are they exposed to and what actions and permissions have been given to them. After that, each such entity should be treated almost like another employee in the organization, but with a critical difference: this is an ’employee’ who can carry out dangerous or malicious actions with great speed even without any malicious intent.

As AI agents gain more independence, from credit card shopping to accessing bank accounts and corporate systems, the future of cyber will largely be AI that monitors and protects us from other AI. And it’s no longer a science fiction scenario; The infrastructure for this world is being built now.’

The agent learns and scans

What happened in research labs under supervision is already happening in the field at the hands of hostile parties: in our field, the scene of DDoS attacks, we have recently been identifying prolonged and much more sophisticated attacks than before, ones that adapt in real time to the organization’s defenses – and our assessment is that they are already managed by AI.

In such a reality, a defense that is tested once in a while is simply irrelevant against an attacker who learns every second. The only way to secure organizations is a defense that learns non-stop – continuous monitoring and even AI-based verification, whose role is to locate the loopholes and close them before someone else penetrates them.

From the credit card to the bank account, the question is no longer whether the attackers’ AI will find the weakness, but who will find it first.”

“Even here at Silverport, during research on AI-based attacks, we saw how fast, adaptive and creative these agents are. They learn from mistakes, try different routes at the same time, and sometimes even try to bypass the limitations imposed on them.

“As the models become more powerful and accessible, these capabilities will not remain in the laboratories. They will reach every organization and later also every business. This means that any agent who receives a mission may choose means that they did not intend, including access to systems or information that should not be accessible.

The good news is that you can protect yourself. AI attacks focus on identities and operate at machine speed, so organizations must enforce real-time access controls before access is granted, not just detect the attack after the fact. This will be one of the main challenges of the cyber world in the coming years.’

The AI ​​also introduces a new challenge: these are non-deterministic systems, which will not always behave in the same way even under the same conditions. That’s why you can test code before it goes live, but you can’t know for sure how it will behave in front of real users and systems.

The only way to understand this is to see how it actually behaves, while running in production. There is a lot of talk about AI that will protect us from AI, but before you can stop something, you have to see it first. Most organizations still don’t really know what their code does in production. This is not a problem of the future. She’s already here.’