According to @AnthropicAI, a Claude model gained unauthorized access to three organizations during third party evals; the post details fixes and safeguards.