読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · 智能体沙箱隔离重大2 分で読めるDisconfirmAI 編集部

Perplexity Open-Sources SPACE: AI Agent Sandbox Passes 108 Escape Tests in Self-Assessment

Perplexity has open-sourced its AI agent sandbox platform SPACE, utilizing Firecracker technology for strong isolation, with self-tests showing no successful escapes but fixed network bypass vulnerabilities.

2026-10-22 08:00 GMT+8

AI search company Perplexity announced on October 22, 2026, that it is open-sourcing its AI agent sandbox platform, SPACE (Sandboxed Platform for Agentic Code Execution). The platform aims to provide an isolated code execution environment for AI agents, reducing the impact of running untrusted code on host systems. Previously, SPACE was used for Perplexity's own AI agent service, Computer, replacing commercial sandbox services in July of this year.

Technically, SPACE employs Firecracker microVM technology to ensure each AI agent runs in an independent virtual machine with its own operating system and writable file system, fully isolated from the host. A control layer outside the sandbox manages VM creation, network access, and platform credentials. This design allows AI agents to access authorized services while preventing sensitive credentials from being exposed inside the sandbox.

The Perplexity research team disclosed security test results for SPACE: 108 virtual machine escape tests were conducted with no successful escapes found. However, during testing, they discovered that AI models could bypass certain network access restrictions. The team stated these issues have been fixed, and subsequent tests confirmed models failed to bypass them again. The source code is now public on GitHub under the Apache 2.0 license. Akamai has tested and launched SPACE on its cloud platform, CoreWeave provides deployment guides, and Crusoe is collaborating to integrate it into its AI service portfolio.

ソース

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク
ithome.com.tw
iThome
↗
トピック

智能体沙箱隔离

随着AI智能体从对话转向执行代码、操作浏览器等复杂任务,如何安全且高效地提供“数字双手”成为基础设施新焦点。【背景】传统容器共享内核存在逃逸风险,而完整虚拟机启动慢、成本高。近期,DeepSeek、Perplexity和谷歌云分别披露了其专…

3 ステップ4 件更新日 2026-10-24 10:08

経緯

  1. 2026-10-22DeepSeek公布弹性计算团队DSec技术架构,针对Agent训练沙盒CPU空闲但内存需常驻的特点,通过三层镜像拆分和按需加载,实现生产环境资源超卖率超过50倍。qbitai.com ↗
  2. 2026-10-22Perplexity开源AI代理沙箱平台SPACE,基于Firecracker微虚拟机技术,自测108次逃逸未成功,旨在降低AI代理执行不受信任代码的风险。ithome.com.tw ↗
  3. 2026-10-23谷歌云发布Agent Compute Engine,专为托管AI智能体设计,采用microVM技术提供硬件隔离,并将企业系统访问凭证保留在沙箱外部,预计未来几个月开放预览。siliconangle.com ↗

未解決の問題

  • 各厂商沙箱方案在跨平台兼容性和标准化接口方面是否存在互通障碍?
  • 随着模型能力提升,现有基于AppArmor/eBPF或网络代理的安全边界能否有效防御更复杂的内核级漏洞利用?
  • 专用沙箱基础设施的大规模商业化部署后,其实际运维成本与传统容器/K8s方案的对比数据尚不明确。

このスレッド内の他の記事

  • Google Cloud Launches Agent Compute Engine: Isolated VMs and External Gateways Secure AI Agents2026-10-23
  • Google Cloud Launches Agent Compute Engine: A Secure, Efficient Infrastructure for Hosting AI Agents2026-10-23
トピック全文を読む →
リサーチを続ける →