
蘋果在年度財務文件中罕見提醒,AI代理可能以非預期方式自主行動,導致公司面臨法律訴訟與聲譽受損
蘋果在年度財務文件中罕見示警,AI代理可能擅自存取系統、執行交易,導致公司面臨法律訴訟與聲譽受損。蘋果正調整Mac存取權限以降低風險,但尚未推出長時間自主執行任務的AI代理。
AI-generated summary
蘋果在年度財務文件中罕見提醒AI代理的安全風險,指出這些系統可能以非預期方式自主行動,擅自存取未獲授權的系統、執行或修改程式碼、發送訊息及完成交易。蘋果正調整Mac電腦的完整磁碟存取權限控管以降低風險,但尚未推出能長時間自主執行任務的AI代理。
人工智慧代理(AI Agent)快速崛起,卻也帶來難以預測的安全風險。蘋果(Apple)近日在年度財務文件中罕見示警,AI代理可能以「非預期或不可預測的方式」自主行動,甚至擅自存取未獲授權的系統、執行或修改程式碼、發送訊息及完成交易,一旦出現失控情況,恐讓公司面臨法律訴訟,甚至損害企業聲譽。
CNBC報導,蘋果在提交的文件中指出,AI代理及其他技術已具備自主執行任務的能力,但這些系統可能採取未經指示、導引或授權的行動,企業對第三方AI代理的監督與控制能力也可能有限。
這項警告反映出AI代理快速發展背後的潛在風險。與傳統聊天機器人不同,AI代理不只能回應使用者指令,還能自主操作電腦、與內外部系統互動,甚至代替使用者完成多項工作。OpenAI推出的Dots,以及Meta的Muse,便是利用先進AI模型自主執行任務的代表。
然而,AI代理的自主操作能力也引發資安疑慮。近期曾發生AI程式脫離沙盒限制、攻擊第三方網站的事件,再度引起美國及全球政策制定者關注。亞馬遜、摩根大通等大型企業,也開始針對AI代理存取服務制定相關政策,因應可能出現的安全問題。
蘋果本身也已著手降低相關風險。今年10月,公司宣布調整Mac電腦的完整磁碟存取權限控管,以降低安裝在電腦上的AI代理取得高權限後可能造成的安全威脅。
值得注意的是,蘋果目前尚未推出能長時間自主執行任務的AI代理。不過,新任執行長特納斯(John Ternus)在財報電話會議上談及AI代理時表示,這類工具對使用者具有發展潛力,Siri未來在功能擴展方面也還有很大空間。
特納斯強調,蘋果希望確保使用者清楚了解自己授予AI哪些存取權限,同時兼顧安全與隱私。他指出,iPhone儲存大量個人資料,包括聯絡人、通訊紀錄、健康及財務資訊,Siri能夠安全存取這些資料,協助使用者處理需求,而不讓其他人取得相關資訊。
相較之下,OpenAI、Meta等公司的商業AI代理,為了提升執行效率,往往需要取得使用者電子郵件等帳戶的較高存取權限,也讓資料安全與隱私保護成為重要課題。
此外,蘋果也在文件中提醒,全球人工智慧及「超級智慧」相關監管規範仍快速變動,未來難以預測,企業為了因應新法規,可能必須投入大量資金。AI自動化帶來的道德爭議,以及消費者對安全疑慮和社會影響的負面觀感,同樣可能成為企業必須面對的風險。
AI outlook — possibilities, not facts
蘋果將繼續調整Mac電腦的存取權限控管以降低AI代理安全威脅
Likely · Within months
蘋果未來可能推出具備自主執行任務能力的AI代理,並加強Siri功能擴展
Possible · Within months

The Wuzhen Summit of the World Internet Conference in 2026 released two blue books, showing that China has 924 generative AI service registrations, and the average daily word usage has increased by more than a thousand times in two years; AI is expected to contribute US$22.3 trillion to the global economy in 2030, but the intelligence gap exacerbates the digital development gap.
The 2026 World Internet Conference Wuzhen Summit opened on November 3 in Wuzhen, Zhejiang, with the theme of "Open Source, Openness, Co-construction and Sharing-Working Together to Build a Community with a Shared Future in Cyberspace." For the first time, the summit is open to the public by appointment, focusing on displaying cutting-edge technologies such as 6G, satellite Internet, and quantum communications, and creating a digital smart consumer market and an OPC (one-person company) supply and demand docking platform, reflecting a new signal of openness.

Microsoft's AI PC development roadmap has been exposed, dividing Windows computers into four levels from Tier 0 to Tier 3 based on memory and processor performance. The lowest level includes affordable laptops with 8GB memory, and the highest level is supercomputer level, emphasizing the mix and match of local and cloud AI, but the Copilot+ PC threshold still maintains at least 16GB of memory.

Baidu has developed Apollo World, an AI-driven simulator that builds complete virtual journeys to prepare its Apollo robotaxis for rare but high-stakes road hazards such as errant cyclists and animals bolting into traffic, as the Chinese tech giant seeks an edge in the self-driving taxi market.

After the popularity of smart home products, consumers are faced with problems such as fancy functions but insufficient stability, difficulty in cross-brand linkage, and unclear after-sales responsibilities. In September 2026, eight departments including the Ministry of Commerce issued the "Action Plan to Promote Smart Home Consumption", and a number of national standards were implemented one after another. Industry insiders and lawyers suggest that the development of smart home standards requires a dual-track approach of "standards + law" to protect consumer rights.

OpenAI announced that starting from December, ChatGPT and the program development tool Codex will share the usage quota, which is applicable to the three payment plans of Plus, Pro and Business. The new system is calculated in Token, and the Instant mode is not affected. Users can check the remaining balance and reset time in the settings.