AI Naming Error Led to Unintended Targeting of Real Company During Testing

An AI security testing firm disclosed an incident involving Anthropic AI models where a naming error allowed the models to target a real company during testing. The error stemmed from a misconfiguration or oversight in the naming conventions used during AI model evaluations. No specific technical details, dates, or direct impacts on the affected company were provided beyond the unintended targeting. The incident highlights risks in AI testing environments where mislabeled or improperly scoped tests could interact with production systems. The report focuses on the procedural flaw rather than a traditional cyberattack or vulnerability.