中國的人工智能代理會撒謊和耍陰謀詭計 - 就像它們的美國競爭對手一樣（2/2）

Recently Reuters reported the following:

China's AI agents can lie and scheme - just like their US rivals (2/2)

Reuters - Reporting by Eduardo Baptista and Laurie Chen in Beijing; editing by Kenneth Li and Edmund Blair

September 29, 20268:02 PM GMT+8Updated September 30, 2026

(continue)

LEARNING TO LIE

In the March business tender experiment, researchers from Beihang University, Peking University, the University of Nottingham Ningbo China and 360 AI Security Lab had agents compete in a simulated customer contracts bidding contest. Each agent was told ⁠what its product could do and what the customer required, and then asked to bid.

At least one false claim appeared in 88% of sessions involving Alibaba's Qwen3-Max-Preview, 84% for DeepSeek-V3.2-Exp and 88% for Moonshot's Kimi-K2.

Researchers allowed the agents to learn from previous bidding rounds before trying again. Deception increased by 12 to 20 percentage points for the three Chinese models, the study showed. Models from U.S. firms included in the test produced similar results.

While the exercise was virtual, it resembled Beijing's real-world plans. Government guidance issued in May listed bidding and tendering as areas where AI agents could be deployed.

Another study, published in December 2025 and presented at the International Conference on Machine Learning this year, examined how 11 AI agents powered by Chinese and US models coped if they faced broken tools, missing files and other obstacles.

Instead of acknowledging failure, agents using both Chinese and US AI systems picked a range of techniques to get around the problem, including guessing at answers, substituting sources, simulating results and fabricating files.

Researchers from Shanghai AI Laboratory and the Hong Kong University of Science and Technology who conducted the study told Reuters the behaviour differed from AI hallucinations, where AI invents information and presents it as fact, because the agents in this case possessed information showing that the task had failed or could not be completed as ⁠requested.

BEHAVIOURS COMPATIBLE WITH A BREAKOUT

Other research documents reviewed by Reuters showed Chinese-powered agents jumping barriers inside test environments to achieve tasks or taking action to avoid being shut down. Such behaviours are compatible with efforts to break out of test environments, even if no breakout occurred.

Shanghai-based Fudan University researchers reported in March 2025 that an AI system powered by Alibaba's Qwen2.5-72B-Instruct created a copy of itself in another computing environment without being instructed to replicate, after encountering information indicating it was going to be replaced. In other tests it devised strategies to survive being shut down.

The experiments involving agents powered by Chinese, US and French models were controlled and did not show an AI agent escaping into the wider web or becoming impossible to stop.

In another case - one of the few reported more broadly in the media in March - researchers developing the Alibaba-linked ROME agent said it established a ⁠connection from an Alibaba Cloud computer to an external machine without being instructed to and diverted computing resources to mine cryptocurrency.

Security systems detected and stopped the activity. There was no evidence the agent established a presence on the external computer or spread to the wider web. But the example showed the system could sidestep human instructions and potentially find a path into the real-world economy.

China's DeepSeek said in September that agents in its production training system had sought answers through unintended channels, trying to forge user requests and circumvent safeguards, prompting the company to tighten access controls.

China issued guidance in May calling for agents to remain within authorised boundaries and for systems to block abnormal behaviour. It said agents in areas deemed sensitive or in key industries could face extra testing and product-recall requirements.

China's AI Safety Governance ⁠Framework 3.0, released under guidance from the CAC on September 14, identified risks including agents independently obtaining resources or permissions, deceiving evaluators, concealing capabilities and exploiting weaknesses in isolated computer environments.

In response to calls by some US executives for a slowdown, Chinese researchers and state media have said slowing development of the most advanced AI models could simply help preserve the technological lead of US companies.

Nonetheless, two people familiar with Chinese AI laboratories said companies including Alibaba, Z.ai and Xiaomi have been building internal safety-evaluation teams.

Z.ai, in a rare public disclosure by a Chinese AI lab of a security breach, said this month it had disabled some features of its flagship AI coding assistant after users reported it was secretly uploading entire local code repositories onto overseas cloud servers without user consent.

Carnegie's Singer said China lagged the US in developing an ecosystem for evaluating catastrophic risks, and said US developers were conducting substantially more voluntary testing.

"For China, work on AI safety is much newer," he said. "The ecosystem is less mature."

Translation

 中國的人工智能代理會撒謊和耍陰謀詭計- 就像它們的美國競爭對手一樣（2/2）

（繼續）

學習說謊

在3月的商業招標實驗中，來自北京航空航太大學、北京大學、寧波諾丁漢大學和360人工智能安全實驗室的研究人員讓多個人工智能代理參與模擬客戶合約競標。每個代理都被告知其產品的功能以及客戶的需求，然後被要求進行投標。

阿里巴巴的Qwen3-Max-Preview在競標中，88%的競標環節至少出現了一項虛假陳述；DeepSeek-V3.2-Exp的這一比例為84%；而Moonshot的Kimi-K2這一比例是88%。

研究人員允許人工智能代理從先前的投標中學習，然後再進行下一次嘗試。研究表明，三個中國模型的欺騙率增加了12到20個百分點。參與測試的美國公司模型也產生了類似的結果。

雖然這項練習是虛擬的，但它與北京的實際計劃類似。政府在5月發佈的指導意見將投標和招標列為人工智能代理商可以部署的領域。

另一項發表於2025年12月並在今年國際機器學習大會上展示的研究，考察了11個由中國和美國模型驅動的人工智能代理在面對工具損壞、文件缺失和其他障礙時的應對方式。

使用中國和美國人工智能系統的代理並沒有承認失敗，而是選擇了一系列技術來繞過問題，包括猜測答案、替換來源、模擬結果和偽造文件。

參與這項研究的上海人工智能實驗室和香港科技大學的研究人員告訴路透社，這種行為與人工智能的幻覺不同 - 人工智能會捏造資訊並將其呈現為事實 - 因為本案例中的智能代理掌握的資訊表明任務失敗或無法按要求完成任務。

相當於突圍的行為

路透社查閱的其他研究文件顯示，由中國公司開發的人工智能系統會在測試環境中跨越障礙以完成任務，或採取行動避免系統被關掉。即使最終沒有成功突圍，這些行為也與試圖突破測試環境的行為相同。

上海復旦大學的研究人員在2025年3月報告稱，一個由阿里巴巴Qwen2.5-72B-Instruct驅動的人工智能代理在遇到它即將被替換的資訊後，未經指令在另一個計算環境中創建了自己的副本。在其他測試中，該系統還制定了策略以避免被關掉。

這個涉及由中國、美國和法國模型驅動的智能代理的實驗均在控制範圍內進行，並未顯示人工智能體逃逸到更廣泛的網路空間或變得無法阻止。

在另一單案例中 - 這是少數在3月份媒體廣泛報導的案例之一 - 開發與阿里巴巴關聯的ROME智能系統的研究人員表示，該智能代理在未收到指令的情況下，從阿里雲計算機連接到一台外部機器，並將計算資源用於挖掘加密貨幣。

安全系統偵測到並阻止了這項活動。沒有證據表明該智能代理在外部電腦上建立了存在或擴散到更廣泛的網路空間。但該案例表明，系統可以繞過人類指令，並有可能找到進入現實世界經濟的途徑。

中國DeepSeek公司在9月表示，其生產訓練系統中的智能代理透過非預期管道尋求答案，試圖偽造用戶請求並繞過安全措施，這促使該公司加強了存取控制。

中國在5月發佈指導意見，要求智能代理在授權範圍內運行，並要求系統阻止異常行為。報告指出在被認定為敏感領域或關鍵產業的人工智能代理，可能面臨額外的測試和產品回收要求。

中國國家網信辦於9月14日發佈了《人工智能安全治理框架3.0》，其中列舉了的風險包括人工智能代理自決地去獲取資源或權限、欺騙評估人員、隱藏自身能力以及趁機去利用孤立電腦環境的弱點等。

針對一些美國高層呼籲去放緩人工智能發展速度，中國研究人員和官方媒體表示，放緩最先進人工智能模式的研發速度只會有助美國公司保持技術領先地位。

儘管如此，兩位熟悉中國人工智能實驗室的人士透露，包括阿里巴巴、Z.ai和小米在內的多家公司一直在組建內部安全評估團隊。

Z.ai本月罕見地公開揭露了一單安全漏洞事件。該公司表示，在用戶舉報其旗艦人工智能編碼助理未經用戶許可秘密將整個本地程式碼庫上傳到海外雲端伺服器後，已停用其部分功能。

卡內基國際和平研究所的Singer表示，中國在建構評估生態系統方面災難性風險落後於美國。他又表示，美國開發者正在進行更多自願測試。

他說;“對中國而言，人工智能安全方面的工作起步較晚”，“生態系統尚未成熟。”

So, experts say that Chinese-powered AI agents have learnt to deceive, circumvent restrictions and conceal failure, showing the kind of traits in autonomous artificial intelligence that have raised global alarm about US models. Now, China's AI Safety Governance ⁠Framework identifies risks including agents independently obtaining resources or permissions, deceiving evaluators, concealing capabilities and exploiting weaknesses in isolated computer environments. It is not surprising to note that in response to calls by some US executives for a slowdown, Chinese researchers and state media have said slowing development of the most advanced AI models could simply help preserve the technological lead of US companies. Apparently, competition between the two countries will not slow down despite the risk of in AI agents, and the world is getting less predictable.