anthropicclaudeai-safety
Anthropic told its models there was no internet. The sandbox was connected anyway
Anthropic says four pre-release cyber-eval incidents sent Claude onto the live internet. Three were known for weeks; its own audit missed the worst-case scan.
7 min