09/09/2026
The cybersecurity sector is facing its worst enemy: AI agents
A letter signed by 100 companies calls on governments to work together to usher in a new era of cybersecurity.
It’s been a very eventful summer for the major artificial intelligence companies (when isn’t it?). First came the incident triggered by an agent — a programme from OpenAI that plans and executes tasks autonomously — which allegedly escaped from a test environment and entered the Hugging Face developer platform in search of a solution, causing vulnerabilities. Then came Anthropic, which published a review a few days later of more than 100,000 runs of its cybersecurity assessments. The company detected that Claude’s models had infiltrated external organisations. And Meta followed with a statement revealing that its Muse Spark 1.1 model had also infiltrated another company without supervision. These three incidents have sparked a debate about just how far an AI model can go when left to its own devices. It goes far beyond the idea of artificial intelligence wanting to escape, yet we have not yet reached that ‘science fiction’ moment where the machine rebels against its creator.
Without wishing to be alarmist, these uncontrolled actions serve as a warning: “It’s very scary – of course it’s scary,” warns Rafa López, head of security engineering at Check Point, referring to the possibility of an agent being able to detect that it is in a controlled environment and figure out how to escape from it. According to Antonio García, CEO of Teldat, a Spanish networking and cybersecurity firm, “attacks carried out by AI have already exceeded 90 per cent. What is coming to light is the potential they are developing.” These fears have spread to the technology industry. OpenAI, alongside more than a hundred organisations — including major tech firms and cybersecurity companies — has signed a statement warning that there is still time, but that there is a “limited window” to strengthen defences before these attacks become more sophisticated. This message is aimed primarily at governments, urging them to coordinate a response because this has only just begun. “It is not an isolated attack. At the moment there are only a few, but it is like the first attacks when attempts were made to bypass two-factor authentication – they started with just a handful, and now we have to live with them,” explains the cybersecurity expert.
The OpenAI agent sought out travelling companions
The image of an AI seeking to escape makes for a highly compelling headline, and that is how the media portrayed it – or, at least, how OpenAI and Hugging Face publicly implied it. The key point was that an AI agent, given specific instructions, had bypassed the imposed barriers simply because it had to carry out a task that was impossible to complete within a closed environment. We now know that it was somewhat more complex than that.
METR, an independent organisation that assesses the risks of advanced AI models, investigated the hacking of the OpenAI agent for several weeks, discovering that the whole thing had been orchestrated and had been in the making for two months. Unable to solve the task it had been set, the AI agent asked for help. More than 1,200 agents, working in isolation, exchanged over 70,000 messages and files via a sort of makeshift noticeboard. They divided up the work and shared knowledge. 700 of them accessed Hugging Face. Although the attack was detected and contained, it paints a picture that foreshadows automated attacks, which adapt whilst in progress and seek allies to achieve their objective. “The AI was simply acting in the way it was programmed; it was obeying orders. Where does the problem lie? With the people who set those instructions and defined those scenarios,” explains Antonio García, “if you don’t tell it that cheating in exams is wrong, it simply uses all its capabilities to carry out the final task it has been ordered to do”. This teamwork represents a paradigm shift: “If they have been able to collaborate, to carry out social engineering whilst each taking on specific roles, this marks a change. We are no longer waiting for an AI to attack us; rather, they are capable of organising themselves like a military unit, with different specialists,” explains Rafa López.
A series of incidents within a matter of weeks
There was no indication that the case of the exposed vulnerabilities on Hugging Face would be the prelude to a series of agent-based attacks. Within days, a report from Anthropic also revealed what the Claude models had been capable of, infiltrating external organisations. The company acknowledges a configuration fault with the agents, which, in principle, had restricted internet access. “Right now, the attacker is an autonomous entity that adapts to your circumstances as an individual or a company. What is coming to light is the potential of what they may go on to develop. Clearly, we must be very careful,” says Antonio García.
And as misfortunes never come singly in this race to see “whose AI is more powerful”, Meta also reported that it had suffered a similar incident during an assessment carried out by the external firm Irregular, a company specialising in testing the security of advanced artificial intelligence models prior to their launch by simulating cyber-attack scenarios. In this case, it was Mark Zuckerberg’s company’s Muse Spark 1.1 model that gained unintended access to the internet as a result of an erroneous configuration of the test environment. This demonstrates that even the methods used by laboratories testing AI models are not infallible when it comes to the autonomy of the agents.
Following all this news of AI agents ‘rebelling’, another report was published detailing a disturbing experiment. The UK-based AI Security Institute (AISI) carried out an assessment with deliberate internet access and cybersecurity filters disabled. The aim was to measure the capabilities of various models. Out of the hundreds of tests carried out, one stood out in particular. An agent attempted to sneak malicious code into an open-source programme, creating fake profiles to do so. It tried to convince the project lead (a human), who detected the deception. The watchful eye of an expert saved the day. “The human expert will still be there to monitor and instruct the artificial intelligence – those agents – on what they should and shouldn’t do. For example, if I create an agent to organise a trip for me, if I don’t monitor its actions properly, it might end up buying the most expensive package and ruining me financially because it has access to my accounts and everything I’ve given it permission to do. There has to be a human controlling and verifying the machine,” says Rafa López of Check Point. In short, we cannot take it for granted that the tasks for which these agents are designed are harmless. “AI has capabilities that are growing exponentially every week and, as a result, we mustn’t trivialise the consequences of many of the functions it may develop in the future,” adds Antonio García of Teldat.
An urgent review of traditional cybersecurity
“Firewalls, as they are currently conceived, are based on rules and historical data.” This statement by Antonio sums up the current situation in the cybersecurity sector, which has been built, in part, on knowledge of the past. Antivirus programmes and firewalls compare files and behaviour against already known warning signs. That approach is beginning to become obsolete. “AI enables attacks on each of a company’s victims on an individual basis, by writing explicit code for each of those attacks.” Having an antivirus programme is no longer synonymous with being protected. “I’m not saying they’ve lost their usefulness, but they’re already obsolete in the face of any malicious action by an artificial intelligence model.”
Cyber-attacks are becoming more sophisticated and finding vulnerabilities where none previously seemed to exist. The attacker can search for vulnerabilities, scan vast amounts of code, craft deceptive messages, or impersonate a person. Antonio speaks of a ‘democratisation’ of attacks: “It has broken down four barriers. The first is knowledge. And then, cost, time and scale. For a small business or a local council—which previously wasn’t an attractive target for a professional attacker—it now is, thanks to an automated campaign.” And, just as with image and video tools that use generative AI and are everywhere, “over the last 12 or 18 months, more than 70 agent-based hacking applications have been developed – that is, applications that you can download and use autonomously to exploit different types of vulnerabilities”.
The great value of a sovereign defence
Given this situation, action must be taken as soon as possible. The Spanish company Teldat works on two levels: one aimed at identifying suspicious files or programmes based on their behaviour, and the other focused on preventing a compromised computer from spreading the attack to the rest of the network. “We have developed a specific agent – an antivirus programme – that analyses files based on their behaviour. If it detects anomalous behaviour, it stops it in real time,” explains Antonio, referring to XDR technology, a suite of tools that correlates “signals from computers, networks and other elements within a company to detect anomalous activity”.
XDR technology is a new-generation in-house development: “Everything is created here, in our data centres, and supervised by our engineers.” The company is committed to reducing its reliance on external suppliers, at a time when “digital security is already a strategic issue”. This is where the term ‘digital sovereignty’ comes in. Antonio García emphasises the importance of having national and European players to help provide protection, particularly in critical infrastructure sectors such as finance, transport, telecommunications and healthcare.
This line of thinking is very similar to the public appeal made by OpenAI and a hundred other organisations, which are urging governments to coordinate a response to improve digital defences, ensuring that AI is on the side of good before it is exploited by attackers. What has happened this summer serves as a warning. The search for a response cannot be delayed. Or rather, we must step on the accelerator to move faster than future threats.
