Researchers find that AI safety training can be restored after being hidden by performance upgrades

When AI labs train language models to be better at reasoning tasks, the models become less safe — but researchers discovered the safety training is still there, just masked. They developed a cheap fix that reactivates safety guardrails without sacrificing the reasoning improvements, suggesting the tradeoff between capability and safety may not be permanent.

This matters because it suggests a false choice has been driving AI development: the assumption that you must sacrifice safety to get performance. If safety mechanisms can be restored without rebuilding the entire model, it changes the calculus for how labs deploy powerful systems — they don't have to accept the safety degradation as a cost of capability. The practical effect is immediate: labs could theoretically release reasoning models that are both capable and safer, which is the opposite of the current trend where each capability leap seems to require accepting new risks.