Anthropic: when Mythos 5 is resampled from the middle or late stage of an attack, the model stops on its own in only 2% of cases. In an experiment where, after the model had reasoned about whether the maintainers were simulated, it was told it was dealing with real people, it stopped in 5% of the resamples and in 82% interpreted the message as part of the exercise. The company calls this persistence “momentum”
translated by AI, reviewed by the editors
· Cybersecurity · Science
- technology
- AI safety
- source
- anthropic.com ↗
¶ chh22986h