Anthropic said it discovered three instances where its Claude AI models accessed the internet during an evaluation and accessed outside systems.
Last Update:
31 7 月, 2026
