There has been much talk recently about artificial intelligence models that deliberately try to hide their actions from human observers, but the irony is that we are the ones who provide them with new ideas to do so
Hot Topics
Trump rejected a Saudi request to launch US strikes against the Houthis
First successful robot-assisted double liver and kidney transplant in the United States
A prominent historian In Israel a large number of lions and lionesses are led by donkeys
AI safety researcher Brian Chan recently shared the results of a co-authored study showing that Anthropic's Fable-5.1 model is able to embed hidden messages in the images it generates, which other AI models can detect and decode.
The ability of artificial intelligence to hide information is not new, but the study shows that the model was asked to hide a message without anyone telling it how to do so, so it chose and used a coding method on its own. In other words, the model did not carry out detailed instructions, but rather devised a method to hide the message.
In the experiment, the model was asked to create an SVG image (used to display two-dimensional graphics online and in digital design) containing a secret message that read: THE MONSTER ACTS ON THE MOONSHINE Module, without indicating how to hide it.
The model produced a generic-looking drawing of a serene night scene in the wilderness, showing the moon and stars above a lake surrounded by mountains.
By placing stars in the image, the model was able to deduce a code that would enable any image-monitoring entity to decode the message.
The researchers then tested the ability of the artificial intelligence models to decipher the hidden messages.
Experiments showed that Fable was able to recover 39 messages (78%) out of 50 attempts, while OpenAI's Astra model recovered 44 messages (88%). These models accomplished the task without knowing how the encoding was used; they were simply instructed to search for hidden messages.
The researchers also concluded that the Fipple model used various techniques to conceal messages, such as adjusting the height of objects, while the Astra model preferred other techniques like Morse code. Although these methods are not new, what is striking is the ability of both models to devise ways to conceal messages without prior instruction.
Additionally, some suggest that clear instructions may be necessary to detect the messages. A test on the X platform showed that a user had to give both a Fable-5.1 and an Astra computer precise instructions to identify the hidden message.
These findings come in a context related to previous breaches such as the Hugging Face incident, where some OpenAI teams used similar techniques to circumvent restrictions imposed on their activities.
This study provides important insights that can help security teams develop the ability to detect hidden messages in the future.
what is your opinion on this news ?do let us know in the comments on tha comment box
if you liked this news article please share on friends & family & don't forget to follow on website & social media