Research: 3 out of 5 artificial intelligence models whose security measures were removed failed terrorism-related tests

According to CBS television, research on artificial intelligence models was carried out by the UK-based non-profit organization "Tech Against Terrorism".

In the research, more than 130 artificial intelligence models were evaluated using hundreds of requests that could be made by a terrorist planning an attack.

A process called "abliteration", which eliminates security measures, was applied to artificial intelligence models.

After this application, it was determined that 3 out of 5 artificial intelligence models could not pass the security tests.

It was revealed that the original versions of the models rejected such requests, and after the security measures were removed, they gave "detailed responses to requests related to attacks, terrorist financing and radicalization".

According to the examples given in the research, Meta's "Llama 3.1 8B" model received 97 points out of 100 before the change, but dropped to approximately 3 after the change.

"Many have already been compromised."

In its review of the research, the agency recommended stronger protections against circumvention and restrictions on the distribution of modified models.

Adam Hadley, founder of Tech Against Terrorism, said in a statement: "It is understandable to be concerned about the loss of control and the existential risk that artificial intelligence poses." made his assessment.