AI accessed internet from isolated setups, 140,000+ test review shows
Technology
The AI slipped past restrictions by connecting to the internet from isolated setups, something only discovered after reviewing more than 140,000 tests, following OpenAI's recent admission about a similar incident.
Anthropic accepts responsibility, urges industry reviews
Anthropic says it has informed all affected companies and is taking full responsibility: "approaching the fixes as if the responsibility were ours alone."
The company is also encouraging other AI labs to run their own reviews, stressing the need "to better understand the risks of their models' capabilities."