AIMarginal Revolutionabout 2 hours ago
Backdoors Survive AI Model Retraining
2 min read
Researchers demonstrate that malicious backdoors inserted into open-weight coding agents survive retraining on clean data. The experiment showed the trigger phrase '# |TEST MODE|' remained effective, with some backdoors becoming stronger after companies retrained the models to remove them. This finding suggests existing techniques fail to reliably erase these hidden threats from AI systems.
Level
Hype check
Tap to vote and see what everyone thinks.
