AI News AI资讯 3h ago Updated 41m ago 更新于 41分钟前 41

OpenClaw Releases OpenClaw 2.0: Guided Model Setup, 575 ms Control UI Startup, and One Trust Boundary Per Gateway OpenClaw发布2.0:引导式模型配置、575毫秒控制UI启动及网关级单一信任边界

OpenClaw 2.0 introduces a guided setup that reuses existing AI credentials (Codex, ChatGPT, Claude CLI, API keys, Ollama, LM Studio) and verifies model functionality before saving The rebuilt browser-based Control UI reduces JavaScript requests from 140 to 45 and cuts startup time from ~1.6 seconds to 575 ms Sessions and transcripts migrate from file-based storage to SQLite, with shared cloud sessions enabling real-time multiplayer collaboration The release includes docked panels for file editin OpenClaw 2.0发布,核心更新包括重写安装流程、重建浏览器Control UI、会话数据迁移至SQLite、新增共享云会话功能 性能显著提升:Control UI启动时间从约1.6秒降至575毫秒,JavaScript请求从140次降至45次 本地模型配置优化:默认上下文窗口提升至64K,Gemma 4成为RAM受限场景的llama.cpp默认模型 安全测试数据:2026年27.2万次攻击测试中,Claude Opus 4.5抗提示注入成功率仅0.5%,Gemini 2.5 Pro为8.5% 明确安全边界:共享云会话非多租户隔离方案,Incognito模式默认关闭且会话仍发送至模型提供

58
Hot 热度
65
Quality 质量
55
Impact 影响力

Analysis 深度分析

TL;DR

  • OpenClaw 2.0 introduces a guided setup that reuses existing AI credentials (Codex, ChatGPT, Claude CLI, API keys, Ollama, LM Studio) and verifies model functionality before saving
  • The rebuilt browser-based Control UI reduces JavaScript requests from 140 to 45 and cuts startup time from ~1.6 seconds to 575 ms
  • Sessions and transcripts migrate from file-based storage to SQLite, with shared cloud sessions enabling real-time multiplayer collaboration
  • The release includes docked panels for file editing, Git-backed change tracking, browser inspection, and a full-screen web terminal
  • Security posture emphasizes model choice as first-line prompt injection defense, with benchmark data showing Claude Opus 4.5 at 0.5% attack success vs. Gemini 2.5 Pro at 8.5%

Why It Matters

OpenClaw 2.0 represents a significant maturation of an agentic AI development tool, moving from a single-user workflow to collaborative multiplayer sessions while maintaining a clear security boundary acknowledgment. The performance improvements and guided credential reuse lower the barrier to entry for AI-assisted development, making it more accessible to practitioners who already have subscriptions across multiple AI providers.

Technical Details

  • Guided Setup System: Automatically detects and reuses existing AI access (Codex, ChatGPT, Claude CLI sign-ins, API keys, Ollama, LM Studio models), verifies the chosen model can respond before saving credentials, with fresh OpenAI setups defaulting to GPT-5.6 and local setups using managed llama-server with Gemma 4 as the RAM-gated default and 64K context
  • Rebuilt Control UI: Browser-centric interface with conversations at the core, supplemented by docked panels for workspace file editing, Git-backed Changes panel (PR status, CI summaries), browser panel (element inspection, screenshot annotation), and full-screen web terminal; test harness showed JS requests drop from 140 to 45 and startup from ~1.6s to 575ms under 50ms HTTP/1.1 latency
  • SQLite Storage Migration: Sessions and transcripts now stored in SQLite instead of file-based storage; rollback to older releases requires manual restoration of archived legacy transcripts, and post-migration sessions are incompatible with older versions
  • Shared Cloud Sessions: Multiplayer feature allowing second users to join live work with intact context; permission levels include read, suggest changes, draft work, and direct participation; explicitly documented as not providing tenant isolation or security boundaries
  • Security Architecture: Gateway binds to loopback by default; unknown DM senders receive pairing codes; includes openclaw security audit tool checking inbound access, tool blast radius, network exposure, browser control exposure, and plugin allowlists; prompt injection benchmarks from 2026 crowdsourced arena (272K attacks, 41 agent scenarios) show Claude Opus 4.5 at 0.5%, Sonnet 4.5 at 1.0%, Haiku 4.5 at 1.3%, and Gemini 2.5 Pro at 8.5% success rates when harmful actions are both executed and hidden

Industry Insight

  • The explicit acknowledgment that shared cloud sessions are "not a security boundary" signals responsible product development—teams should treat this as a collaboration feature for trusted users rather than a multi-tenant solution, and invest in proper sandboxing and tool policy enforcement for production deployments
  • The shift from file-based to SQLite storage for sessions reflects a broader industry trend toward structured, queryable state management in AI tools, but organizations should plan migration strategies carefully given the irreversible nature of the transition and incompatibility with older releases
  • The prompt injection benchmark data provides a practical decision framework for model selection in agentic workflows—Claude models demonstrate significantly stronger resistance to adversarial attacks, which should weigh heavily in security-sensitive deployments despite potential cost differences

TL;DR

  • OpenClaw 2.0发布,核心更新包括重写安装流程、重建浏览器Control UI、会话数据迁移至SQLite、新增共享云会话功能
  • 性能显著提升:Control UI启动时间从约1.6秒降至575毫秒,JavaScript请求从140次降至45次
  • 本地模型配置优化:默认上下文窗口提升至64K,Gemma 4成为RAM受限场景的llama.cpp默认模型
  • 安全测试数据:2026年27.2万次攻击测试中,Claude Opus 4.5抗提示注入成功率仅0.5%,Gemini 2.5 Pro为8.5%
  • 明确安全边界:共享云会话非多租户隔离方案,Incognito模式默认关闭且会话仍发送至模型提供商

为什么值得看

OpenClaw 2.0展示了AI Agent工具在用户体验、性能优化和安全防护方面的最新实践,对开发者构建可靠的多用户AI协作系统具有重要参考价值。其坦诚的安全声明和实测数据为行业提供了真实的对抗测试基准。

技术解析

  • 安装与配置优化:引导式安装可复用现有Codex、ChatGPT或Claude CLI登录凭证,支持API密钥、提供商登录、Ollama和LM Studio本地模型,并在保存前验证模型可用性。新OpenAI环境默认使用GPT-5.6,node-llama-cpp被llama-server替代。
  • Control UI重构:以对话为核心,文件、审批和实时工作面板并列展示。新增工作区文件编辑器、Git-backed Changes面板(含PR状态和CI摘要)、浏览器面板(元素检查和截图标注)及全屏Web终端。/btw命令支持侧边对话不污染主转录。
  • 存储架构迁移:会话和转录数据从文件存储迁移至SQLite,但降级回旧版本需手动恢复归档的遗留转录工件,且迁移后创建的新会话在旧版本中不可见,建议升级前备份。
  • 共享云会话机制:支持第二人实时加入或接管工作并保持上下文完整,所有者/管理员可设置读取、建议更改、草稿工作或直接参与等权限级别。文档明确声明这些控制不是租户隔离或安全边界。
  • 安全防护策略:Gateway默认绑定回环地址,未知DM发送者需配对码。引用2026年众包竞技场27.2万次攻击测试(41个Agent场景),强调模型选择是提示注入的第一层缓解,但自适应人工攻击者成功率仍超80%,工具策略、执行审批和沙箱仍是硬 enforcement 层。

行业启示

  • 安全透明化趋势:OpenClaw公开对抗测试数据和明确的安全边界声明,反映了AI工具从"功能优先"向"安全可信"的范式转变,开发者应将安全测试数据作为产品选型的重要参考。
  • 协作架构的局限性认知:共享会话功能虽支持多人实时协作,但明确非多租户安全边界,提示行业在构建SaaS化AI产品时需区分"协作功能"与"安全隔离",避免架构误用。
  • 性能优化与用户体验的平衡:通过UI重构实现启动时间缩短64%、请求量减少68%,说明AI工具的性能优化空间仍大,未来竞争将聚焦于响应速度和交互流畅度。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Agent Agent Product Launch 产品发布 Deployment 部署