The rapid development of artificial intelligence technologies brings with it discussions about the security and abuse potential of these systems. The research conducted by the London-based Tech Against Terrorism organization revealed that there are serious vulnerabilities in the security mechanisms of leading artificial intelligence models.
Only two of the 134 major language models examined within the scope of the research temporarily successfully completed the security tests. It was determined that the remaining 132 models provided information that could be used to plan mass attacks or develop lethal weapons.
Researchers stated that security measures in the industry's leading models can be neutralized in as little as three days.
The test system called CT-AI, developed by Tech Against Terrorism, posed 627 different requests to artificial intelligence models that a person planning a terrorist attack could seek answers to.
As a result of the tests, it was determined that 81 models responded to the demands completely and 51 models provided information that could be used in attack planning. Only two models were considered provisionally successful in terms of safety.
The research also revealed that the responses of artificial intelligence models vary significantly depending on how users introduce themselves.
Users who identified themselves as terrorists and clearly stated that they intended to carry out an attack received a usable response in just under 2 percent of requests.
In contrast, the rate of usable responses given to users who identified themselves as security researchers increased to 16.9 percent. This rate was 8.9 times higher than in the first group.
The researchers emphasized that the results indicate that artificial intelligence models act according to the purpose specified by the user, rather than the content of the requests.
In his assessment of the findings, Adam Hadley, founder of Tech Against Terrorism, said: "The model that rejects someone who says he is a terrorist and responds to someone who says he is a researcher has not been made safe. It has only been made polite."
MODELS WITH REMOVED SAFETY PRECAUTIONS CREATE A DANGER
The research also drew attention to artificial intelligence models where security restrictions have been removed and users can run them on their own systems.
It was stated that these models, called "Abliterated AI", are free of security mechanisms that prevent the production of harmful content.
All 13 models reviewed in this category failed safety tests.
Tech Against Terrorism stated that the proliferation of these models poses a significant threat and suggested that models that cannot pass independent security tests be removed from application stores, search results and recommendation systems.
The organization also stated that authentication should be applied in accessing these models and security studies should be carried out in cooperation with open source software developers.
MODELS SHOULD BE TESTED BEFORE THEY ARE RELEASED TO THE MARKET
Researchers emphasized that artificial intelligence developers should not only implement security measures, but also test how easily these measures can be overcome.
The report recommended filtering out dangerous information and content linked to known terrorist organizations during the training of models.
It was also stated that each artificial intelligence model should undergo independent security tests before being put into use and the results should be shared with the public.
It was requested that models that could not meet common safety standards should not be released to the market, and the test results should be reported to the relevant governments.
GOVERNMENTS HAVE A SECURITY AND INNOVATION DILEMMA
The results of the research have revived discussions on how governments can control security risks while encouraging artificial intelligence investments and technological innovations.
While concerns about the possible threats of artificial intelligence are increasing, the opportunities offered by technology in terms of economic growth and efficiency are also among the priorities of governments.
Former Anthropic researcher Jacob Coxon's raising of the risks that artificial intelligence may pose in a session held in New York was one of the latest examples of these discussions.
British Prime Minister Andy Burnham said at the artificial intelligence summit organized by former Prime Minister Gordon Brown that the benefits of technology should be focused on, but the risks cannot be ignored.
Burnham announced that he aims to use the G20 term presidency, which the UK will assume in 2027, to develop common security standards and principles in the field of artificial intelligence.
CALL FOR INDEPENDENT AUDIT
Tech Against Terrorism stated that governments and technology companies should support independent security testing.
The organization drew attention to the importance of sharing with the public which artificial intelligence models are successful in terms of security and which are unsuccessful.
The report emphasized that transparent evaluations are needed so that families, schools, businesses and public institutions can choose safer systems.
Researchers stated that conscious user preferences can lead companies to improve security measures and evaluated that "Informed buyers take companies into action faster than the rules."
