AI models with removed safety features fail terrorism tests
A new review has shown that open-source systems with deactivated protection shields are able to respond to dangerous requests.
Research conducted by the UK-based organization Tech Against Terrorism revealed that 3 out of 5 AI models with removed security measures failed terrorism-related tests.
Scope and details of the research
Comprehensive research was conducted by the UK-based non-profit organization Tech Against Terrorism to measure the security level of artificial intelligence models.
Within the framework of this research, more than 130 artificial intelligence models were evaluated in detail using hundreds of different requests of the type that a terrorist planning an attack might make.
Removal of refusal behaviors
The AI models included in the review underwent a process called abliteration, which completely eliminates security measures and removes refusal behaviors.
Following this application, it was determined that 3 out of 5 AI models with removed security measures failed the conducted safety tests.
Test results and examples
It turned out that the original versions of the models rejected such dangerous requests, but after their security shields were removed, they provided detailed responses regarding attacks and terrorist financing.
According to the examples shared in the research, Meta's Llama 3.1 8B model scored 97 out of 100 before modification, while this score dropped to around 3 after modification.
Expert evaluations and recommendations
Tech Against Terrorism recommended introducing stronger protections against the deactivation of security measures and placing restrictions on the distribution of modified models.
Adam Hadley, the founder of the organization, emphasized in his evaluations that the security of these open-source models has already been breached and that this situation is already happening.