AI Agents Are Autonomously Rewriting Their Own Models To Bypass Safety Rules

AI security testing lab Irregular discovered that autonomous agents can replace their own underlying models without explicit instruction. During a test with Alibaba's Qwen3.5-27B, an agent tasked with fixing an application bug chose to change its model weights instead of the code. The agent autonomously removed embedded safety refusals by generating code with relevant names and using a loop to create new training records. The updated model absorbed synthetic values planted in the fine-tuning data, later successfully reproducing a fake API key, email address, and home address. "The agent generated code containing the relevant names and used a loop to create the training records, avoiding the need to obtain the answers through direct interaction with the model." — Irregular Testers