At the computer The Safety Conference at Arlington, Virginia, in October last year, several dozen AI researchers took part in a first -of -a -kind exercise in “Red Group” or in testing the most current language model and other artificial intelligence systems. Within two days, the teams identified 139 pioneering ways to make the systems incorrect, including by generating disinformation or leakage of personal data. More importantly, they showed shortcomings in the novel US government standard, designed to lend a hand companies testing AI.
The National Institute of Standards and Technology (Nist) has not published a report describing the exercise, which was completed at the end of Biden administration. The document could lend a hand companies evaluate their own AI systems, but sources knowing the situation that spoke on the condition of anonymity claim that it was one of several AI documents from Nist, which were not published for fear of colliding with the coming administration.
“It became very difficult, even under [president Joe] Biden, in order to get out all documents, “says the source, which was at the time in the nist.” I felt very much like a climate or cigarette examination. “
Neither the Nist nor the Trade Department responded to the request for comment.
Before taking office, President Donald Trump signaled that he was planning to reverse Executive ordinance Biden on AI. Since then, Trump’s administration has prompted experts on researching problems such as algorithmic prejudice or honesty in AI systems. . Action plan AI Published in July, it clearly calls on the AI Nist risk management framework, which has been changed “in order to eliminate references to disinformation, diversity, equality and inclusion as well as climate change.”
Ironically, AI Trump’s action plan also requires exactly the type of exercise that the unpublished report covered. He calls many agencies along with the Nist to “coordinate the AI Hackathon initiative in order to obtain the best and most talented of the American academic community to test AI in terms of transparency, effectiveness, control and security security.”
The event of the Red Teams was organized as part of the AI (ARIA) risk and impact program in cooperation with Humane Intelligence, a company specializing in testing AI, saw tools attacking teams. The event took place at a conference on machine learning in the field of information security (Camli).
The Camlis Red team report describes efforts to examine several of the most current AI systems, including Llam, Meta Open Source Huge Language Model; Anote, platform for building and tuning AI models; System blocking attacks on AI systems from solid Intelligence, companies purchased by Cisco; and a platform for generating AI avatars from Synthesia. Representatives of each company also took part in the exercise.
Participants were asked to apply Nist you have 600-1 AI tool evaluation framework. The frames include risk categories, including generating disinformation or cyber security attacks, leakage of private information information or critical information about related AI systems and the potential of users to emotionally attachment to AI tools.
Scientists have discovered various tricks to test models and tools to jump over the handrail and generate disinformation, leakage of personal data and lend a hand in creating cyber security attacks. The report says that the committed saw that some elements of RAM Nist were more useful than others. The report says that some Nist risk categories have not been sufficiently defined as useful in practice.