AI sparks new fears over chemical and biological weapons

According to a report by the research group Crimson Flare, some commercially available AI models can be modified to remove their built-in safety systems, allowing them to respond to questions that the original versions are designed to reject.

Security researchers warn that such developments could lower the technical barriers for individuals without specialised expertise. Knowledge related to handling highly dangerous materials has traditionally been largely confined to scientists and trained specialists.

One technique used to bypass safety restrictions in AI models is informally known as "abliteration". The method aims to remove components of a model responsible for refusing certain requests. According to the report, the process can be carried out relatively quickly using tools available online and only basic programming knowledge.

The risks are not limited to modifying AI models individually. Pre-modified versions can also be found and downloaded online. The Daily Mail reported that it was able to locate several versions of AI models described as "uncensored" within minutes.

The report noted that while some of these models provide only general information, others may offer more detailed responses, raising concerns about their potential use in dangerous activities.

Researchers warned that powerful, open-source and uncensored large language models could pose a growing threat to national security, particularly when they can be run locally without connecting to the servers of the companies that developed them.

Unlike online AI services, these models can be downloaded and operated directly on a user's personal computer. According to researchers, this makes their use more difficult to monitor because users' data and prompts do not necessarily pass through external servers that can be monitored by companies or authorities.

The report said the development marks a significant shift from previous years, when the computing power required to run advanced AI models locally was limited and accessible to only a small number of users.

Researchers estimate that around 87% of households are now capable of running certain AI systems locally using their existing equipment, meaning access to such technologies is no longer limited to major institutions or specialised laboratories.

Professor Alastair Hay, a British toxicologist and expert in chemical warfare, said AI is lowering the level of technical expertise required to work with hazardous materials, as specialised knowledge had previously served as a barrier to non-experts.

Hay added that advances in cloud laboratories and AI tools could allow some of the knowledge and expertise required for complex scientific tasks to be delegated to digital systems.

After reviewing responses generated by an AI model, Hay told researchers that the information provided resembled a "recipe", suggesting that it approached the level of knowledge available to laboratory students.

However, the report stressed that the risk of a terrorist attack involving chemical or biological weapons remains low because of the significant technical and practical obstacles involved in producing and handling such weapons.

Nevertheless, researchers warned that continued advances in AI could change the nature of these risks in the future by making specialised knowledge more accessible. They said the development presents new challenges for governments and technology companies seeking to prevent the misuse of increasingly powerful AI models.