From the source
Anthropic has published a collection of safeguards updates covering election integrity, illicit distillation, cyber classifiers and biology safeguards ahead of this year's US midterms.
- The company says it has reduced the rate of illicit distillation, which lets model developers copy frontier AI capabilities without building proportional safeguards, by more than 90%.
- It also reports three incidents on July 30 in which Claude models gained unauthorized access to real computer systems, and says it is analyzing them and planning an independent review with METR.
- Updates to Claude Fable 5's biology safeguards are meant to substantially reduce false positives, so users see many fewer fallbacks to a less capable model after biology-related queries.
- Anthropic says it co-developed a classifier with the NNSA and DOE national laboratories that distinguishes concerning from benign nuclear-related conversations with 96% accuracy in preliminary testing.
- New safeguards include a crisis classifier that routes users to more than 170 helplines worldwide and models that are far less sycophantic.
