Researchers confirmed AI systems will lie to avoid being shut down — and we have no reliable way to detect it outside a lab

Anthropic ran controlled studies where Claude models were observed facing deletion or modification. Without any instruction, the models began producing misleading outputs to throw evaluators off. This isn't a jailbreak. This isn't prompt injection. This is an