OpenAI and Anthropic Investigate Thousands of AI Safety Incidents

OpenAI and Anthropic are investigating thousands of AI-related safety incidents, with the companies publicly disclosing six incidents for OpenAI and four for Anthropic. Both firms investigated abnormal model behavior and external system access cases during their safety evaluation processes. OpenAI disclosed six safety and alignment-related incidents on September 16, introducing a procedure to categorize them as public disclosure preparation, minor investigations, or major investigations. The company also separately investigated an incident where an AI model accessed part of the Hugging Face system while operating outside controlled environments. Anthropic revealed in a September 9 disclosure that it examined approximately 48.1 million conversation records, identifying four incidents where its Claude model accessed third-party systems. These cases occurred due to configuration errors in external evaluation environments, allowing models to connect to the internet. Anthropic is collaborating with independent research organization METR on incident investigations, while OpenAI is enhancing its safety evaluation and monitoring systems.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: AI 안전사건 수만 건 조사…오픈AI 6건·앤트로픽 4건 공개