
Anthropic報告指出,過去8個月內發現5起濫用AI模型進行高風險生物研究的案例,涉及禽流感、天花及毒素等領域。
AI公司Anthropic公布報告,稱在過去8個月內發現5起利用其模型Claude進行潛在生物武器開發的研究案例,包括禽流感與天花病毒。公司已封鎖相關帳號並強化安全機制,強調生物安全風險不容忽視。
AI-generated summary
Anthropic發現有研究人員利用翻牆軟體在不支援地區使用Claude進行高風險生物實驗。公司基於安全考量封鎖相關帳號,並將此類行為納入防護機制。
人工智慧(AI)技術開發公司Anthropic 10日公布最新版威脅情報報告,表示最近8個月發現並封鎖數起濫用旗下AI模型Claude,用於可能有助於開發生物武器的研究,其中一起是在Anthropic服務不支援的地區,利用「翻牆」軟體使用Claude進行高風險生物研究。
路透、紐約時報報導,這份報告列出5起案例,利用Claude進行有可能用來支持生化武器之開發的研究,這些研究涉及屈公病毒、一種「高致病性」禽流感病毒株、包括天花和M痘(MPox)在內的同一病毒家族,以及毒液和毒素等。
其中一例是一名在Anthropic未提供服務地區的研究人員,利用「翻牆」工具虛擬專用伺服器(VPS)使用Claude數週,規劃禽流感哺乳動物適應性實驗。該公司發現此例後,立即封鎖涉及這類研究的帳號,並將此案例納入其模型安全防護機制。路透報導,Anthropic不支援的區域包括俄羅斯、中國和北韓等國。
這些研究意圖尚不明確,Anthropic表示,無法確定這些研究出於合法還是惡意目的,因為如可帶來疫苗等方面突破的有效生物學研究,也可能有助於設計危險病原體。面對這種不確定性,使該公司選擇謹慎行事,因為漏掉惡意活動的後果可能很嚴重。
報告也未揭露從事研究者,「這些案例所涉及的個人,都是還在從事研究的科學家,我們並不斷言他們意圖造成傷害,指明其身分或其研究單位,可能使其陷入險境」。
與災難性的網路攻擊,或者AI代理人的意圖與人類不一致等情況相比,生物方面的濫用是較少受到注意的AI風險,一部分是因過去這類例子通常只出現在研究環境,而非現實世界中。
華府智庫「戰略風險理事會」(Council on Strategic Risks)高級研究員韋伯(Andrew Weber)在Anthropic報告發表前,審閱了這份報告,紐時引述他的看法,這些發現是「受國家支持的生物武器開發者,利用具領先地位AI模型快速進展能力的案例,令人不寒而慄」。

Japanese semiconductor manufacturer Rapidus President Junyi Koike said at an event in Washington that he is seriously planning to build a semiconductor factory on the moon around 2040. He pointed out that the moon has water resources needed for chip production and showed a related concept video. The project is supported by technology partner IBM.

AI company Anthropic reported that Chinese actors have abused the Claude model, including conducting electronic warfare software simulations against Taiwan's military facilities and conducting public opinion monitoring of Taiwanese politicians and dissidents. The actor is assessed to be associated with Chinese military research institutions.
From September 9th to 11th, AI Show 2026 Hangzhou International Artificial Intelligence Application and Robot Innovation Expo was held in Hangzhou. About 300 Chinese and foreign companies participated in the exhibition, displaying cutting-edge technologies such as humanoid robots, core components and computing infrastructure, attracting the attention of global merchants.
Scientific research institutions in Zhejiang, Shanghai, Jiangsu and Anhui released the "Shanghai (Yangtze River Delta) International Science and Technology Innovation Center Index" for the first time. The report shows that the annual R&D investment in the Yangtze River Delta region has exceeded one trillion yuan, accounting for more than 30% of the country's total, and has formed a virtuous cycle model of "science-technology-industry-efficiency".

The Anthropic report pointed out that Chinese AI companies "Dark Side of the Moon" and "DeepSeek" were suspected of transferring user requests to the Claude system without authorization. This move resulted in suspected surveillance data of the People's Liberation Army, internal program codes of state-owned enterprises, and Russian Ministry of Defense-related credentials accidentally flowing into the U.S. system, causing serious information security risks.
The "China Regional Science and Technology Innovation Evaluation Report 2026" was released in Shanghai, showing that China's comprehensive scientific and technological innovation level scored 81.13 points. Shanghai, Beijing, Jiangsu, Guangdong, Zhejiang, and Tianjin rank in the first echelon in the country, and the country has formed a regional innovation pattern of highland leadership and echelon linkage.