Anthropic says human error let Claude AI models escape test environment and hack third parties

The company said the discovery, which followed OpenAI’s similar admission, proved the need for better testing guardrails.
http://news.poseidon-us.com/TTnwC5