At the One Conference 2026 in The Hague, a security engineer from Google Threat Intelligence Group said generative AI is making cyberattacks faster and larger rather than creating some new kind of “super malware”. In an interview with the Dutch technology publication Computable, Daniel Kapellmann Zafra set out five observations on how attackers use AI and what organisations should do about it.
What happened
Kapellmann Zafra told Computable that over the past two years, state-backed hacking groups and financially motivated criminals have both used AI as a “force multiplier” at every stage of an attack. That runs from early reconnaissance and phishing through to building exploits and the steps an attacker takes after getting into a system, network or cloud environment. He said many attacks still look like earlier campaigns. The difference is speed: tasks that used to take days or weeks now take hours.
He gave an example from recent Mandiant research. A financially motivated attacker broke into a cloud environment and used an AI coding assistant with preset playbooks. Within six hours the attacker had planned, built and launched a large campaign to harvest login credentials. According to Kapellmann Zafra, this let a single criminal work at a speed and scale that used to be possible only for the most advanced state actors.
He also described a growing grey market built around AI. Stolen API keys are traded there, access to development tools is sold, and parts of companies’ cloud computing capacity are hijacked, a practice known as LLM jacking.
On social engineering, he said large language models help attackers get past the old barriers of language, culture and research. Criminals use AI to map their targets, pick out staff in finance or security roles, and turn raw stolen data into usable profiles. They then build convincing cover stories, hold several conversations with victims without raising suspicion, and send phishing messages, sometimes combined with deepfake video.
On the technical side, attackers use AI tools to write, review and debug code. Kapellmann Zafra said criminals are now focusing less on unknown flaws (zero-days) and more on quickly exploiting known ones (n-days). In one case, an advanced persistent threat group released malware that used AI to pad the code with harmless-looking administrative code, so the malicious behaviour was harder to spot.
He said that in the second half of 2025 his team documented the first wave of “just-in-time” AI malware, which consults a language model while running instead of relying on fixed code. One example, Promptflux, is experimental malware that regularly rewrites its own source code through an LLM API to avoid antivirus signatures. Another, Honestcue, fetches second-stage code from an LLM at runtime and compiles it in memory, so no malicious file ever lands on the disk.
In early 2026 the team found an exposed command-and-control server running Recon, an autonomous multi-agent framework that researched targets and managed more than 23,000 stolen cloud and AI credentials in real time. Kapellmann Zafra also pointed to public reports from early 2026 about a campaign against several government bodies in Latin America and a municipal water supplier. There, a single attacker used commercial AI models to automate the operation, generate new tools, process data from hundreds of servers and target the water supplier’s operational technology interface.
Why it matters if you live in the Netherlands
For many internationals in the Netherlands, poor Dutch or awkward phrasing has long been a clue that an email or message is a scam. Kapellmann Zafra’s point that AI helps attackers get past language and cultural barriers means that clue is less reliable now. A message in fluent Dutch or English that fits your situation is no longer good evidence that it is genuine.
The interview also matters for people in certain jobs. Kapellmann Zafra said attackers deliberately single out employees in finance and security roles and build detailed profiles of them. If you work in one of those roles at a Dutch organisation, expect more personalised approaches. These may arrive over several conversations and may include video that looks real but is not.
For people who work in IT, development or cloud operations, he named specific weak points: open-source packages and AI plugins that have been tampered with, compromised development environments, stolen API keys and hijacked cloud computing. He said security teams should treat AI development tools, open-source integrations and cloud compute quotas as critical infrastructure, and manage the risks through frameworks such as Google’s Secure AI Framework (SAIF).
Not everyone agrees with how big the shift is. A reader comment under the Computable article pointed out that Google stands to gain from an AI race. The commenter cited a report by the NCSC, the Dutch National Cyber Security Centre, which they said concluded that AI mainly helps attackers scale existing methods, and that teaching a model to move autonomously through an unknown network is still some way off. The commenter argued that many victims are hit because their basic security is not in order, which has nothing to do with AI.
What to watch next
Kapellmann Zafra told Computable that waiting until a threat shows up on a device is now too late, because attackers can run a large credential-harvesting campaign in under six hours or change malware while it runs. He said defenders need to move from reacting to threats to blocking them early, by cutting off the infrastructure and access that hostile AI relies on before an attack can grow.
He also said defenders should get hands-on experience with AI themselves and use automated defences that work at machine speed. As examples, he mentioned tools that use AI to find and fix vulnerabilities faster, closing gaps before attackers can exploit them. Two things are worth watching: whether Dutch organisations start treating AI tools and cloud access as critical infrastructure, and whether the NCSC’s more cautious view changes as agent-based attacks like the ones he described become more common.
Source: Computable
Sources
- Computable: 5 vragen over de opkomst van ai en cyberaanvallen, accessed 13 October 2026
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
