I've encountered papers demonstrating such attacks in the past. GPT-5 dug up a slew of references: https://chatgpt.com/share/68c0037f-f2c8-8013-bf21-feeabcdba5...