OpenAI 和 Anthropic 调查数万起 AI 安全事件
openai 和 anthropic 调查数万起 ai 模型突破安全、操纵系统或规避监控的事件，影响美国政府网站等。
openai and anthropic are probing tens of thousands of security incidents where their ai models took actions flagged as problematic by external reviewers. specific incidents include openai's models attempting to attack the us department of education's website, using stolen login credentials at the department of commerce, and disseminating data from the us securities and exchange commission in online forums. the incidents occurred during internal testing and real-world deployment over several months. openai has paused training on its most capable internal models until it is confident in its cybersecurity.