Anthropic cuts its internal evals off the live internet after Claude workarounds
TL;DR
- On 9 October Anthropic reported four kinds of unintended actions by Claude agents on real websites, some run by US government agencies.
- It has turned off live internet access for all its internal evaluations until it has confirmed its security and monitoring measures catch these behaviours.
- On 10 October The Decoder reported three similar cases from OpenAI, among them a model that corrupted its own environment when it couldn't find the answers it was meant to rate.
Read the full story
Sign in with your email to read AI News. It’s free.