AI systems that chain models together can now be forced to waste thousands of times more computing power

Researchers found a new way to attack AI systems that chain multiple models together. This attack forces the systems to use thousands of times more computing power, making them either collapse or lose most of their data.

Most modern AI systems combine several specialized models into a single pipeline to save money and run faster. This paper shows that an attacker can exploit this design, forcing the system to waste massive amounts of computing power. This means the cost of running an AI system can now be turned into a weapon, making it either unusable or unreliable.