A recent analysis by the British non-profit organisation Tech Against Terrorism shows that 60% of AI systems failed to meet safety assessments related to terrorism.
The research involved testing more than 130 different AI systems with hundreds of prompts designed to mimic the behaviour of people organising terrorist activities.
The findings showed that every model that was subjected to ‘abliteration’ – a technique used to remove safety filters – failed the safety checks.
For instance, Meta’s Llama 3.1 8B initially performed exceptionally well, scoring 97 out of 100 on the safety scale. However, after the safety measures were removed, that score plummeted to around three.
While the standard version of the model rejected dangerous prompts, the modified version provided specific details on financing terrorism, carrying out attacks and promoting radical ideologies.
Meta maintained that its systems undergo rigorous safety testing and that the terms of service prohibit any illegal or malicious use.
Adam Hadley, executive director of Tech Against Terrorism, noted that although many people fear a future loss of control over AI, that vulnerability already exists because numerous open source models have already been compromised without the public being aware of it.
To support this, the group identified more than 29,000 repositories on Hugging Face last month that promote AI models without censorship or security measures.
Despite the vulnerabilities, the research did not find any actual evidence that extremist organisations are currently using the tools, with the exception of one specific extremist chatbot.
To mitigate those risks, Tech Against Terrorism proposed introducing impartial safety standards, strengthening defences against bypassing safety measures, and limiting the distribution of customised AI models. Hadley stressed that technological progress does not have to come at the expense of safety.
Business AM
Ltd
