Anthropic Warns AI Safety Tests Are Becoming Too Easy
As Advanced Models Outgrow Benchmarks, Lab Calls for More Realistic, Multi-Turn Evaluations
As artificial intelligence models become more capable, frontier labs have focused on measuring models' ability to follow directions and remain safe. But the safety evaluations have a problem, according to Anthropic. They're too easy.
As artificial intelligence models become more capable, frontier labs have focused on measuring models' ability to follow directions and remain safe. But the safety evaluations have a problem, according to Anthropic. They're too easy.