Preprint demonstrates attack that gives AI requests higher scheduling priority

Researchers report an attack that manipulates predicted response lengths to give requests higher priority in AI serving systems. In experiments, adversarial requests completed up to 1.53 times faster on average. The attack, called JIL, adds an optimized suffix that makes a lightweight length predictor underestimate the response. Tests using the TRAIL scheduler covered two datasets and four language models, reducing predicted output lengths by up to 83.4 percent-substantially more than the change in actual response length. Response quality varied across models and tasks. Grouping length predictions into coarse intervals reduced the attack's scheduling advantage and mitigated delays to benign requests.