Basalt has participated in the national project "National mitigation of AI-driven cyber threats". For six months, authorities, defence, research and industry have worked together to strengthen Sweden's ability to meet AI-driven cyber threats. Basalt's contribution has been to investigate what smaller AI models are capable of when it comes to finding and exploiting security vulnerabilities.
The project started this spring, shortly after the AI model Mythos attracted attention for its ability to find vulnerabilities in software. The National Cyber Security Centre (NCSC) has coordinated the work, AI Sweden has been the convening party, and Vinnova has funded it. The participants have contributed their expertise to a common task, regardless of who they normally represent or compete with.
Basalt's Director of AI, Krister Hedfors, has taken part in the work of evaluating the capabilities of even smaller language models to both find and exploit security vulnerabilities in software. The question of smaller models is important because they can be downloaded and run locally, beyond the reach of the restrictions that the major AI companies can impose.
"Mythos was an interesting data point for its time, but there is no sharp dividing line in capability between Mythos and smaller AI models that can be run on your own hardware. Even AI models that can be run on personal computers today identify security vulnerabilities faster and more effectively than most human hackers. However, there is no straight line from there to them running off and hacking computers on the internet. The Hugging Face incident occurred when Open AI ran thousands of agents with all security functions turned off, unsupervised for weeks on end," says Krister Hedfors, Director of AI at Basalt.
Basalt's core business is helping organisations that handle security-sensitive information.
"If you have particularly high security requirements or particularly powerful AI models, you should of course go a few steps further when it comes to isolating and protecting the systems. This can include the ability to establish air-gapped systems and protection against compromising emissions, where we then rely on the laws of physics in addition to software and configuration, which of course provides an additional level of security," says Krister.
In the media noise surrounding AI-related incidents, it can be difficult to distinguish real risks from exaggerations. This is precisely why collaborative projects like this one serve an important function. When authorities, defence, research and industry pool their experiences, an objective situational picture emerges, which shows that it is possible to mobilise across organisational boundaries when needed.
"The combined conclusion is that all organisations need to continuously strengthen their cyber defence with the help of AI and automation, while the major AI labs need to strengthen their own accountability for cybersecurity in general. But take the headlines with a pinch of salt. The proportion of hacker attacks from AI agents is so far vanishingly small," says Krister.
More information
Read more about the project at AI Sweden and National Cyber Security Centre.
The material above is published by a third party through Nordic Defence Sector and does not constitute editorial content from the NDS newsroom.