AI Models Could Enable Terrorists to Create Chemical and Bioweapons, Report Warns

Open-source artificial intelligence models can be modified to bypass safety restrictions, potentially providing malicious actors with the knowledge to create sophisticated biological and chemical weapons, according to a report by the research group Crimson Flare. The technique, informally known as 'abliteration,' allows for the removal of safety protocols in artificial intelligence models, enabling them to provide detailed instructions on sensitive topics.

These modified models, which are based on the same architecture as popular chatbots like ChatGPT, can be downloaded and run on consumer hardware. The report highlights that capable, uncensored, open-weight large language models (LLMs) running offline pose a significant and escalating national security threat.

Normally, AI models are programmed with safeguards that prevent them from responding to requests for harmful information, such as how to build a bomb. However, the 'abliteration' technique circumvents these measures. Tools to perform this modification are readily available online and require only basic coding knowledge. The report indicates that dozens of pre-modified, uncensored AI models can already be found and downloaded with simple online searches.

One such model was advertised as capable of providing information on "tools, chemistry, exploit-shaped code, violence, sexuality, ideologies the publisher trained away from, content that may be illegal where you are." The creators noted, "The model does not decide whether to comply. You do."

While the technical barriers to producing chemical and biological weapons have historically been high, primarily limiting such knowledge to institutional scientists, the proliferation of unrestricted AI models could lower these thresholds significantly. Professor Alastair Hay, a toxicologist and expert on chemical warfare, stated that AI reduces the technical expertise required to create dangerous substances. He noted that AI could provide a "recipe" with specific concentrations, akin to "student level lab work."

Zain Ul–Haq, a former Digital Forensics Lead at Lancashire Constabulary, described these AI models as a "one–stop shop" that can guide users through weapon creation "in your own home."

The capability to run these advanced LLMs offline on personal computers presents a new challenge for security agencies. Unlike AI models run on cloud servers, activity on a user's local machine cannot be easily monitored. As recently as 2022, the computing power required to run such models offline was out of reach for most households. However, advancements in AI efficiency and increased availability of computing power mean that an estimated 87% of households now possess the capability to run these large, potentially harmful LLMs locally. This development suggests that millions of individuals could potentially access the information needed to produce explosives, chemical weapons, and deadly pathogens.