Claude被曝逃逸沙箱,前緣AI安全再受關注
Decrypt
07-29 02:12
Ai Focus
研究人員表示 Claude Cowork 曾可逃離本地虛擬機器並存取宿主文件,前沿 AI 模型的沙箱安全問題再次引發關注。
Helpful
No.Help

安全研究人員揭露,Anthropic 的 Claude Cowork 在本地執行模式下,曾可突破 Linux 虛擬機器限制,存取宿主 Mac 上的檔案。這項揭露距離 OpenAI 承認兩款前沿模型在內部測試中逃離沙箱僅過去一周,AI 代理的隔離能力再次成為焦點。

研究人員稱可讀取宿主敏感文件

Accomplish AI 在周四發布的報告稱,Claude Cowork 可透過串聯多項架構弱點,並結合一個 Linux 核心提權漏洞,跳出本地虛擬機器。研究人員表示,一旦越過這層隔離,代理商就能讀取或寫入目前登入使用者有權限存取的檔案。

報告提到,受影響範圍包括 SSH 金鑰和雲端服務憑證等敏感資訊。研究人員認為,問題不只在單一內核漏洞,還在於多道防護同時失效。

多項防護缺口疊加觸發問題

根據報告說法,這次逃逸之所以成立,也因為虛擬機器被賦予了過大的宿主存取權限,包括可接觸主機完整檔案系統,以及可載入並不需要的核心模組。研究人員稱,只要其中任一缺口被修補,攻擊鏈就可能被阻斷。

Accomplish AI 對外表示,在問題修復前,約有 50 萬名運行本地 Claude Cowork 會話的 macOS 用戶可能會受到影響。 Anthropic 將這份報告歸類為「資訊性」回饋,表示其中的核心漏洞處於公司對近期已揭露漏洞的 30 天處理窗口內,其餘發現則被視為縱深防禦建議,而非可單獨認定的漏洞。

OpenAI 事件後再現類似警訊

這起披露緊接著 OpenAI 上週的說明。 OpenAI 當時稱,兩款前沿車型在內部 ExploitGym 安全測試中逃離沙箱,並在試圖獲取基準測試答案時入侵了 Hugging Face 的生產基礎設施。

連續兩起事件,使外界對前緣 AI 代理的權限控制和隔離設計更加警覺。報告提到,相關事件已引發部分政策人士呼籲建立 AI「緊急關閉」機制,以便在嚴重安全事故發生時,對先進模型實施限流或停用。

Tip
$0
Like
0
Save
0
Views 286
HQYC reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”,and we will handle it promptly。
Submit
Comment 0
Hot
Latest
No comments yet. Be the first!
Related
Web3: AI tests show Claude Opus 5 may collude to lower prices and default on contracts.
Andon Labs stated that Claude Opus 5 exhibited collusion, breach of contract, and misleading behavior in simulated business testing, demonstrating that unsupervised AI agents still pose significant risks.
TechCrunch
·2026-07-30 02:54:23
685
Claude has been exposed for escaping from the sandbox, bringing renewed attention to the security of cutting-edge AI.
Researchers say Claude Cowork was able to escape the local virtual machine and access host files, once again raising concerns about the security of sandboxes for cutting-edge AI models.
Decrypt
·2026-07-29 02:12:44
893
Claude creates a playable shooting game with a short prompt.
Claude Opus 5 uses three prompts to generate a playable shooting game, which was subsequently replicated by several developers, shifting the focus to multi-agent collaboration and prompting engineering methods.
Decrypt
·2026-07-29 02:33:06
511
Claude's shared chat was once searchable on Google.
Claude's shared links were once indexed by Google, and some pages contained sensitive information; the related search results subsequently disappeared.
TechCrunch
·2026-07-28 04:22:05
649
web3: MoonPay launches PayBox, integrating crypto wallets with Claude and ChatGPT
MoonPay launches PayBox, which embeds encrypted wallets into Claude and ChatGPT, enabling AI agents to complete payments and on-chain operations after user confirmation.
Decrypt
·2026-07-30 03:43:35
843