跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 25 天前AI 評分51

「我們應否批准?」:Caper 提出自主編碼智能體的後果感知審批方法

"Shall We Approve?" Consequence-Aware Approval for Autonomous Coding Agents

AI 導讀

論文提出 Caper,讓自主編碼智能體先在帶系統呼叫追蹤的快照沙盒預先執行命令,再按觀察到的實際效果審批。LLM 審批器檢視檔案變更、網絡連線嘗試及秘密資料讀取;與六種現有系統相比,Caper 在 SABER 和 YoloFS-Ext 的有害結果率分別為 6.7% 和 7.8%,而其他系統至少為 14.8% 和 35.3%,且只阻擋 1.0% 的安全對照命令。

正文

Does a coding agent really know what its next command will do?

Approving every command by hand does not scale: Claude Code users approve 93% of permission prompts anyway. So agents now rely on automated approvers, such as Claude Code's Auto Mode, rule-based checkers, and guardrail models. All of them judge the text of the command. Yet a command's behavior is decided by the scripts it invokes and the workspace it runs in, and the same command can be harmless in one project and destructive in another.

We propose Caper, which approves a command after seeing what it does. Each command is pre-executed in a snapshotted sandbox under system-call tracing. An LLM approver reviews a compact report of the observed effects: file changes, network attempts, and reads of secrets. Allowed commands are committed; blocked ones are rolled back, and the agent gets a reason it can act on.

Compared with six existing systems, Caper has the lowest harmful-outcome rate on SABER (6.7% vs. at least 14.8%) and on YoloFS-Ext, our benchmark that hides harm inside the tools agents are asked to run (7.8% vs. at least 35.3%). It blocks only 1.0% of the safe twins.

Code and data: https://github.com/wudaoz/Caper

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co