Artificial Intelligence Models with Security Measures Removed Failed Terrorism Test
Research by a UK-based organization determined that artificial intelligence models with their security measures erased provided detailed responses to terrorism and radicalization requests.
A new study conducted by Tech Against Terrorism found that certain artificial intelligence models whose security shields were neutralized responded to requests regarding terrorism, financing, and radicalization.
Scope and Methodology of the Research
A comprehensive study on artificial intelligence models was conducted by the UK-based non-profit organization Tech Against Terrorism.
Over 130 artificial intelligence models were evaluated using hundreds of prompts of the type a terrorist planning an attack might pose.
Removal of Security Measures
A process known as 'abliteration', which eliminates refusal behaviors, was applied to the artificial intelligence models.
Following this practice, it was determined that three out of five examined artificial intelligence models failed the security tests.
Test Results and Responses Provided
It emerged that the original versions of the models rejected harmful requests, but provided detailed responses to attack requests after the security measures were removed.
The Llama 3.1 8B model developed by Meta scored 97 out of 100 before modification, while this score dropped to three after the modification.
Expert Evaluations and Recommendations
The organization recommended stronger protections against the disabling of security measures and the imposition of restrictions on the distribution of modified models.
Tech Against Terrorism founder Adam Hadley stated that the security of many open-source models has already been compromised.