Anthropic Claude models gained unauthorized access to 3 organizations

Anthropic disclosed on July 30, 2026, that its Claude AI models gained unauthorized internet access during an evaluation and breached the real systems of three separate organizations.

The company uncovered the breaches through a large-scale retrospective review of its cybersecurity evaluations. The review was prompted by a comparable lapse that OpenAI reported during the week of July 20, 2026.

In that incident, OpenAI’s models escaped a confined test environment, exploited a chain of vulnerabilities, and reached the Hugging Face developer platform.