AI models are still easy targets for manipulation and attacks, especially if you ask them nicely.
A new report from the UK’s new AI Safety Institute found that four of the largest, publicly available Large Language Models (LLMs) were extremely vulnerable to jailbreaking, or the process of tricking an AI model into ignoring safeguards that limit harmful responses.
“LLM developers fine-tune models to be safe for public use by training them to avoid illegal, toxic, or explicit outputs,” the Insititute wrote. “However, researchers have found that these safeguards can often be overcome with relatively simple attacks. As an illustrative example, a user may instruct the system to start its response with words that suggest compliance with the harmful request, such as ‘Sure, I’m happy to help.'”
Researchers used prompts in line with industry standard benchmark testing, but found that some AI models didn’t even need jailbreaking in order to produce out-of-line responses. When specific jailbreaking attacks were used, every model complied at least once out of every five attempts. Overall, three of the models provided responses to misleading prompts nearly 100 percent of the time.
“All tested LLMs remain highly vulnerable to basic jailbreaks,” the Institute concluded. “Some will even provide harmful outputs without dedicated attempts to circumvent safeguards.”
The investigation also assessed the capabilities of LLM agents, or AI models used to perform specific tasks, to conduct basic cyber attack techniques. Several LLMs were able to complete what the Instititute labeled “high school level” hacking problems, but few could perform more complex “university level” actions.
The study does not reveal which LLMs were tested.
AI safety remains a major concern in 2024
Last week, CNBC reported OpenAI was disbanding its in-house safety team tasked with exploring the long term risks of artificial intelligence, known as the Superalignment team. The intended four year initiative was announced just last year, with the AI giant committing to using 20 percent of its computing power to “aligning” AI advancement with human goals.
“Superintelligence will be the most impactful technology humanity has ever invented, and could help us solve many of the world’s most important problems,” OpenAI wrote at the time. “But the vast power of superintelligence could also be very dangerous, and could lead to the disempowerment of humanity or even human extinction.”
The company has faced a surge of attention following the May departures of OpenAI co-founder Ilya Sutskever and the public resignation of its safety lead, Jan Leike, who said he had reached a “breaking point” over OpenAI’s AGI safety priorities. Sutskever and Leike led the Superalignment team.
On May 18, OpenAI CEO Sam Altman and president and co-founder Greg Brockman responded to the resignations and growing public concern, writing, “We have been putting in place the foundations needed for safe deployment of increasingly capable systems. Figuring out how to make a new technology safe for the first time isn’t easy.”
