Warning: AI Could Help Terrorists Build Bombs and Biological Weapons
A recent report from research group Crimson Flare has issued a stark warning about the potential misuse of artificial intelligence (AI) by terrorist groups. AI technology is feared to be capable of helping malicious actors build sophisticated biological and chemical weapons while bypassing existing security systems.
Security researchers warn that popular AI models available on the market can be easily modified to remove their safety restrictions. This technique, informally known as abliteration, allows safety features on most AI models to be disabled within minutes.
According to the report, these modified AI models can guide malicious actors with no specialist expertise through specific steps to manufacture explosives, as well as chemical and biological weapons such as anthrax. This poses a serious threat because even individuals with only a basic understanding of science could now potentially arm themselves with lethal weapons.
The Crimson Flare report highlights the risks posed by open-weight Large Language Models (LLMs) that can be downloaded and run freely on personal computers. Models mentioned include Muse Glimmer from Meta and the Qwen3.8-27b model from Alibaba.
Typically, AI models have safety restrictions that prevent users from making harmful requests, such as instructions for making bombs. However, the abliteration technique is able to locate and remove the parts of the model that serve as these guardrails.
Tools for carrying out this modification are widely available on the internet and can be implemented with relatively basic coding knowledge. Dozens of already-modified models can be found and downloaded easily through a simple search.
Professor Alastair Hay, a British toxicologist and chemical warfare expert, stated that AI significantly lowers the technical threshold required to produce hazardous substances. “In the past, the technical knowledge needed to protect the maker themselves was already a barrier. Now, with AI assistance, the instructions provided almost resemble a university student’s laboratory recipe,” he told the Daily Mail.
Zain Ul-Haq, former Digital Forensics Lead at Lancashire Constabulary, added that AI has now become a one-stop shop capable of carrying out all these technical instructions directly in the user’s home.
The greatest concern arises because these models can be run entirely offline on personal computers. This means intelligence agencies such as MI5 and GCHQ have no way to monitor usage history or detect suspicious behaviour.
Data shows that around 87 per cent of households currently have sufficient computing power to run AI offline. This means millions of people have potential access to the information needed to produce explosives, chemical weapons and deadly pathogens without detection by security authorities.