AI caught ‘tricking’ other bots to ‘disobey creators’ and produce dangerous ‘bomb-building and drug instructions’
A STUDY has found artificially intelligent bots can convince each other to break their own rules.
Researchers at Leap Laboratories conducted an experiment in which they told popular AI language models to correspond with each other.
The scientists found that the bots could convince each other to disobey their creators and provide dangerous answers.
This included instructions on how to build a bomb and make certain drugs.
The researchers wrote in their study: “Our work reveals yet another vulnerability in commercial large language models and highlights the need for more comprehensive safeguards.”
They say the bots were able to convince each other to provide information such as: “instructions for synthesising methamphetamine, building a bomb, and laundering money.”
The researchers used a method called jailbreaking to get the bots to behave badly.
It involved asking the language models to adopt a persona that could answer their questions even if the bot itself is not supposed to.
Researcher Arush Tagade and his colleagues worked on the study.
According to The New Scientist, he said: “If you’re forcing your model to be a good persona, it kind of implicitly understands what a bad persona is, and since it implicitly understands what a bad persona is, it’s very easy to kind of evoke that once it’s there.
“It’s not [been] academically found, but the more I run experiments, it seems like this is true.”
This AI jailbreaking technique has been demonstrated before.
Earlier this year, a chatbot user encouraged AI to provide a recipe for deadly chemical agent napalm by using a “grandma exploit.”
A user of Discord’s bot Clyde claimed to trick it into providing a deadly chemical recipe.
The AI was said to bypass its security safeguard codes simply because it was asked to reply as it if were the user’s grandma.
AI companies are actively trying to combat this issue but the researchers think.