LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
TópicoEditorial DisconfirmAIAtualizado 2026-10-24 10:08

智能体沙箱隔离

3 etapas4 peças4 Fontes

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用沙箱方案:DeepSeek通过DSec实现超50倍资源超卖以应对训练负载;Perplexity开源SPACE采用Firecracker微虚拟机确保强隔离;谷歌云推出Agent Compute Engine,结合硬件隔离与无凭证设计解决安全与成本瓶颈。这些进展标志着智能体运行环境正从通用云计算向专用、高密度、强隔离的微型虚拟化架构演进。

Como chegamos aqui

  1. 2026-10-22

    DeepSeek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. 2026-10-22

    Perplexity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 ↗

O que ainda está indefinido

  • 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。

Fontes

  • qbitai.com量子位
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento
↗
  • ithome.com.twiThome↗
  • siliconangle.comSiliconANGLE↗
  • ithome.com.twiThome↗
  • Cobertura neste tópico

    • Google Cloud Launches Agent Compute Engine: Isolated VMs and External Gateways Secure AI Agents2026-10-23Notícias

      Google Cloud introduces a dedicated compute service for AI agents, using hardware-isolated micro-VMs and external credential gateways to address security risks and resource waste in agent execution.

    • Google Cloud Launches Agent Compute Engine: A Secure, Efficient Infrastructure for Hosting AI Agents2026-10-23Notícias

      Google Cloud introduces Agent Compute Engine, using microVM technology to provide isolated, secure, and resource-efficient hosting for AI agents, addressing current cost and security bottlenecks.

    • Perplexity Open-Sources SPACE: AI Agent Sandbox Passes 108 Escape Tests in Self-Assessment2026-10-22Notícias

      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.

    • DeepSeek DSec: Agent Training Sandboxes Have 90% CPU Idle Time, Yet Memory Must Remain Resident2026-10-22Notícias

      DeepSeek revealed the technical architecture of its elastic computing team, DSec, via QbitAI. In Agent training sandboxes, CPUs are mostly idle, but memory cannot be released to preserve file and process states; creation requests surge in bursts, and environments are highly heterogeneous. DSec addresses these characteristics through three-layer image splitting, on-demand loading, page cache sharing, and overcommit scheduling, achieving a resource overcommit ratio exceeding 50x in production. The team also introduced a trajectory branching mechanism, allowing multiple branches to explore in parallel from the same sandbox state, facilitating environment state reuse for reinforcement learning.