Researchers challenge hidden AI safety features as unreliably secure

Original: Position: Retire the "Positive Backdoor" Label -- Secret Alignment Requires Strict and Systematic Evaluation

Source: arxiv.org ↗