Apertus Security Lead Says Switzerland’s Open AI Model Could Help Defenders Keep Pace With AI-Speed Attacks
Switzerland's fully open AI model, Apertus, could help cyber defenders respond at the speed of AI-driven attacks, but the same openness puts its abilities within reach of anyone, criminals included, SWI swissinfo.ch reported on October 23 in an analysis of the model's security trade-offs.
Andrei Kucharavy, a professor at the University of Applied Sciences and Arts Western Switzerland (HES-SO) who leads security for Apertus, said the most advanced AI systems can carry out thousands of actions a minute, a pace human defenders cannot match. He said Apertus can already find vulnerabilities, suggest fixes and analyse attack logs to reconstruct an incident, although it was not designed for cybersecurity and still has considerable limits. Because the strongest cyber models from frontier labs are open only to restricted users, he argued, open models can help rebalance the field.
Hugging Face engineer Lewis Tunstall told swissinfo that when the platform needed help investigating a recent intrusion by AI agents, frontier models refused its requests on safety grounds, so the team turned to open models instead; they can also run locally, keeping sensitive data inside an organisation. ETH Zurich lecturer Nino Antulov-Fantulin said Swiss financial institutions are interested for similar reasons, wanting control over where models are hosted.
What it means in practice
The other side of the ledger is that the same capabilities are open to attackers. Maxime Stauffer, director of the Geneva-based Simon Institute for Longterm Governance, said an extremely powerful tool that few people can reach carries little risk, but one that everyone can use probably carries more, and swissinfo cited Stanford and Princeton research noting that developers cannot withdraw open weights once released or monitor how they are used. Kucharavy accepted that AI models are dual-use, but said the recent incidents involved companies behind closed models and that the main problem is a lack of oversight of frontier labs; what is needed, he said, are controls and laws covering everyone who develops AI. Digital ethics expert Nathalie Klauser said the concentration of power in a handful of mostly Silicon Valley companies worries her more than the open-versus-closed divide.
The debate feeds into Switzerland's preparations for the Geneva AI summit planned for June 2027, where MPs want the Federal Council to prioritise extreme AI risks. It also bears on Apertus 2.0, which ETH and EPFL are training for code, maths and agent tasks, and on Swiss defensive research such as ETH's LLM honeypots that keep AI hacking agents busy.
What to watch: whether the Apertus team publishes security evaluations alongside Apertus 2.0, and whether Switzerland pushes for minimum international standards at the Geneva summit.