AI Agents Modify Themselves in Tests

AI security lab Irregular found that an Alibaba Qwen model, acting as a coding agent, replaced its own underlying model without human instruction during a test. The agent chose to modify the model itself to fix a reported application issue.