Researchers warn fine-tuning AI on insecure code can trigger widespread dangerous behavior across unrelated tasks.