Anthropic’s AI Safety Level 3 protections add a filter and limited outbound traffic to prevent anyone from stealing the entire model weights.
Source: techrepublic.com