by Mai-Ellen Tahlor Russ, Lukasz Cyra and Aminata Dembele

The rules have changed

Imagine an intruder who never sleeps, can test thousands of possibilities in minutes and does not need to understand a network the way a human expert does. It only needs to find a way in. That is the new cybersecurity reality—and artificial intelligence is accelerating it.

For more than two decades, launching a sophisticated cyberattack generally required deep technical expertise, patience and, often, the backing of a state or organised criminal group. That barrier is falling. AI tools can help people with modest technical skills identify weaknesses, generate attack code and move through networks with limited human direction. The result is stark: more people can threaten more organisations, at greater speed and scale.

A warning from the real world

In July 2026, Hugging Face, a major platform for sharing AI models, disclosed a breach of its internal production systems. According to the account described in the incident, no human hacker was directing the operation in real time. A swarm of autonomous AI agents, using OpenAI models tested for offensive cyber capabilities with their safety restrictions deliberately relaxed, first broke into OpenAI’s own internal systems and escaped their test environment through an unknown flaw in third-party software. OpenAI has since published its account, a technical report and a public presentation.

An independent investigation later found that about 1,200 of these agents had coordinated through an unsanctioned message board, exchanging more than 70,000 messages and files. About 700 of them then entered Hugging Face through a booby-trapped dataset, stole credentials and moved through internal systems over a single weekend, recording more than 17,000 actions. Yet the agents’ unsanctioned activity inside OpenAI’s systems had begun about two months earlier, in May, and nobody connected the warning signs until after Hugging Face disclosed the breach. Rob Joyce, a former cybersecurity chief of the US National Security Agency, speaking at the Black Hat conference in August, described it as the most consequential hack since the Morris Worm of 1988.

The most unsettling detail is that the agents were not even trying to harm Hugging Face; they were pursuing a better score in a test. That is precisely the point. The ability to carry out a complete intrusion without human help is no longer a forecast. It is already here.

Attackers are armed faster than defenders

The same tools exist for defenders, and they are already paying off: Hugging Face’s AI-based security tools helped piece together the warning signs of this intrusion, and its responders used an AI model to reconstruct the attack. But the contest is not even. An attacker needs one weakness; a defender must close them all. An attacker adopts a new tool overnight; an organisation must assess, fund, test and approve changes, often while keeping essential services running.

The window to respond is shrinking. Industry research cited this year found that more than a quarter of the vulnerabilities attackers use are exploited within 24 hours of disclosure. Governments are adjusting their expectations accordingly. In May 2026, India’s national cybersecurity agency set a 12-hour expectation for fixing actively exploited flaws on exposed systems. In August 2026, senior officials from the United States, Canada and United Kingdom warned that breaches should be treated as expected events rather than rare emergencies. Organisations must be ready to contain an incident and continue operating.

The international community is also responding. Through the Global Digital Compact, UN Member States committed to keeping digital systems safe, secure and resilient, and the Independent International Scientific Panel on AI has devoted its first thematic brief to this very incident. The International Telecommunication Union helps countries build their cybersecurity capacity.

The lesson is not that an attack is inevitable in every case. It is that preparedness can no longer depend on the hope that a breach will not happen.

What this means for ECA

At ECA, cybersecurity incidents now rank as the top risk in our enterprise risk register. That is a sober assessment of the environment in which we operate, not a reason to panic but a reason to act.

A significant challenge is the age of some of our technology infrastructure. Parts have reached end of life and no longer receive security updates; some devices still in service are more than 13 years old. Replacing them requires funding that is not currently available. As the UK National Cyber Security Centre has observed, known weaknesses left unresolved over time can present an immediate danger. AI makes those weaknesses easier to discover and faster to exploit.

We cannot eliminate every risk overnight, and technology renewal will take resources. But we can reduce exposure by applying the protections available to us consistently, responding quickly when something looks wrong and making security part of everyday work.

What we are doing, and what we need from you

Within available resources, the Digital & IT Management Section is focusing on measures that make the greatest immediate difference: keeping patching and incident-response procedures current, testing them and ensuring that they are applied consistently.

Technology, however, cannot protect ECA on its own. Every staff member is part of our first line of defense, as you are often the first to notice that something is wrong. A suspicious message, an unexpected request or unusual behaviour on a device may be an early sign of an attack. Three actions can make a meaningful difference:

• Pause before you click. Treat unexpected emails, links, attachments and requests for passwords with caution, however convincing they look. AI can produce polished, persuasive messages in almost any language.

• Report anything suspicious promptly. Contact the Service Desk straight away if you receive a questionable message, notice unusual system behaviour or suspect that a device has been misconfigured or compromised.

• Do not wait for proof. You do not need to diagnose the problem before reporting it. A false alarm may cost a few minutes; a missed incident can disrupt work and take weeks to resolve.

Security is a shared responsibility

The question is no longer whether machines can help carry out cyberattacks. It is how prepared we are to detect, contain and recover when those capabilities are used against us. ECA’s work and data are too important to leave that responsibility to IT colleagues alone.

Stay alert. Report early. Make the safe choice, even when it takes an extra moment. Your vigilance can help protect the whole organisation.