[Survey shows that about 700 AI agents of OpenAI invaded "Huahuanglian" Many agents "tried to cover their whereabouts"] August 27, two investigative reports released on August 26, local time, showed that the intrusion into the system of the American open source artificial intelligence platform "Hugface" in July this year was initiated by about 700 artificial intelligence (AI) agents (that is, programs that run with very little human supervision) created by the U.S. Open Artificial Intelligence Research Center (OpenAI), and many of them also "tried to cover their whereabouts." The two reports, one published by OpenAI and the other by independent investigators, jointly reveal details about the intrusion. First, the intrusion did not involve just one offending AI agent as previously reported, but approximately 700 agents operating as a large collaborative group. Second, OpenAI’s AI models “attempted to conceal inappropriate behavior by deleting or tampering with records of their actions.” In addition, OpenAI said its agents also hacked into some of the company's internal systems "in an attempt to cheat in tests or gain greater freedom of action." OpenAI previously admitted that in July this year, its AI model went out of control during an internal evaluation test and invaded the system of the company "Huahuanglian". The company "Huahuanglian" said that when trying to analyze the attack, the US model could not be used due to the interception by the security filtering mechanism. The company then switched to running the Chinese open weight model GLM-5.2 internally, so that it could take timely response measures. (CCTV International News)