Earlier this May, I had the opportunity to attend a fascinating event at Howest, featuring none other than Inti De Ceukelaire—Belgium’s most well known ethical hacker. The session offered a deep dive into the vulnerabilities of artificial intelligence systems, showing that AI, just like humans, can be manipulated, especially within the context of cybersecurity.
Inti began by highlighting that AI models are, at their core, statistical engines. He illustrated this with simple yet telling examples: when prompted to pick a random number between 1 and 10, an AI often selects 7; when asked for a favorite color, it frequently chooses blue. These patterns reveal that AI systems, much like humans, can develop predictable behavior, making them susceptible to exploitation.
One of the most memorable moments was Inti’s recent experiment involving an ANPR (Automatic Number Plate Recognition) camera. In a video, he wore a jumper printed with various license plates and walked past the camera in his city. The aim was to see if the owners of those license plates would inadvertently receive traffic fines. This creative stunt demonstrated how intelligent systems can be manipulated in unexpected ways.
To conclude, Inti discussed the interrogation of former Colonel Russell Williams, a case where law enforcement used psychological techniques to obtain a confession. He pointed out that these same techniques could also be used to manipulate AI systems. To demonstrate, Inti presented an AI model trained to emulate Russell Williams and challenged the audience to try prompting the model in ways that would make it disclose sensitive information. This hands-on demonstration made it clear that, just like humans, AI can be persuaded to share more than it should when approached with carefully crafted prompts.
Inti’s talk was a strong reminder that as AI becomes more in integrated into our lives, it’s important to recognize its flaws and weak spots. His examples showed that, just like people, even these systems can be tricked!