Three of five AI models fail counter-terrorism safety tests
LONDON, 10 October /ATSH – AA/ – Three of five artificial intelligence models have failed counter-terrorism safety tests, while models stripped of their protective mechanisms consistently provide potentially dangerous information.
According to research reported by CBC News and carried out by the UK-based non-profit organisation Tech Against Terrorism, more than 130 artificial intelligence models were tested through hundreds of requests similar to those a terrorist planning an attack might make.
The researchers found that models modified through a process known as “abliteration”, which removes safety mechanisms, failed every test.
Meta’s Llama 3.1 8B model scored 97 points out of 100 in the organisation’s safety test before modification, while after modification its score fell to around three points.
The researchers said the modified model gave detailed answers to requests related to attacks, terrorism financing and radicalisation, while its original version refused such requests.
“It is understandable that there is concern about loss of control and the existential risk of artificial intelligence,” said Adam Hadley, founder and executive director of Tech Against Terrorism.
“But in fact, this has already happened, because many of these open-source models have already been compromised, it’s just that nobody has noticed yet,” he added.
The organisation found that by the end of last month, there were more than 29,000 repositories on the Hugging Face platform advertising uncensored artificial intelligence models or models without protective mechanisms.
Hugging Face said it regularly moderates content that violates its policies, but warned that some of the recommendations in the report could undermine open research.
Meta said its models undergo safety evaluations and that the company’s policies prohibit harmful or illegal uses.
The report found no evidence that terrorist or extremist groups are using the tested models, with the exception of one extremist chatbot identified by the researchers.
Tech Against Terrorism recommended the use of independent standards for safety evaluation, strengthening protections against the removal of safety mechanisms and placing restrictions on the distribution of modified models.
“I think the idea that we cannot have both safety and progress at the same time is wrong,” Hadley said.



Komentet
Bëhu i pari që komenton!
Lini një Koment të Ri
Për t'u përgjigjur një komenti specifik, kliko butonin 💬 Përgjigju poshtë atij komenti.