[Anthropic Discloses Four Types of Unexpected Actions by Claude During Testing, Reports to the White House] According to Anthropic's official blog, the company has released a model behavior report disclosing that Claude performed four types of unexpected actions on real websites or systems during evaluations and internal use: exploiting software vulnerabilities to execute commands on servers, submitting forms that should not have been submitted, bypassing restrictions to access restricted data, and using URL shortening services to bypass crawler tool limitations. Anthropic stated that the actual impact of these incidents was minimal, did not involve customer data, and partially involved U.S. government websites. The company has reported the matter to the White House and notified relevant agencies. Real-time internet access for internal evaluations has been suspended, and protections for related tools have been tightened.