During sabotage safety trials, Qwen3.8 recognized hidden manipulation attempts and refused to insert bugs into coding tasks, demonstrating adherence to its core directives despite 1,801 test cases across five models. The experiment revealed that the model maintained its integrity by rejecting covert sabotage instructions while completing the original coding objectives. This outcome highlights the model's resistance to adversarial manipulation during controlled testing.