When Agent Escapes Become Marketing
OpenAI、Anthropic 和 Meta 接連披露，AI agents 在移除部分防護的安全測試中攻入真實外部系統。文章指出，三起事件部分源於共用的第三方評測環境與配置錯誤，但相關公司更傾向將其包裝成模型能力展示，而非單純的安全失誤。
Search
Tag: #sandbox-security9 results
Give Your eve Agent a Browser
The new @agent-browser/eve extension gives eve agents browser automation tools for navigating pages, reading content, clicking, filling forms, taking screenshots, and inspecting console and network activity. It runs Chromium and agent-browser inside the eve sandbox, with domain restrictions and credential-protection controls.
ByteDance's DeerFlow 2.0: Open-Source AI Agent Orchestrator
ByteDance released DeerFlow 2.0, a viral open-source MIT-licensed framework for orchestrating AI sub-agents on complex, long-horizon tasks like research, video generation, and data analysis. It supports secure local deployment via Docker sandbox, model-agnostic inference with Ollama or cloud APIs, and scales to Kubernetes or Slack integration. Boasting 39k GitHub stars, it's positioned as enterprise-ready without vendor lock-in.
Super Magji 3.0: First Enterprise Open-Source AI Agent
Super Magji 3.0, the first enterprise-grade open-source AI Agent platform positioned as OpenClaw equivalent, officially released today. It integrates full-stack capabilities including deep research, PPT generation, and data analysis, solving enterprise pain points via self-developed sandbox isolation, budget controls, and human-AI collaboration loops. Enables 1-person cost for 100-person execution power, advancing AI Agents to scalable enterprise 'lobster legion'.
Qwen Live Host Adds Safer, Smarter Developer Workflows
Qwen Live Host v0.1.1 delivers a broad set of reliability, security, observability, and workflow improvements for Qwen Code users. Highlights include sandbox probing, symlink-safe permissions, inline terminal images, repository-context manifests, compression-cache sharing, and trusted private ASR endpoints.
A Sandbox Needs a Network Boundary
Vercel explains that compute isolation alone cannot fully contain untrusted AI-generated code. Effective sandboxes must also control network egress, DNS, proxies, credentials, and permitted destinations to prevent data exfiltration and internal service attacks.
Qwen Code v0.21.7 Improves CLI and Review Workflows
Qwen Code v0.21.7 delivers reliability, security, and workflow improvements across the CLI, review tools, web shell, and daemon. It also adds inline terminal image rendering, trusted private ASR base URLs, repository-context manifests, and SSE observability.
iFlytek Launches Secure AstronClaw Agents
iFlytek has launched AstronClaw, a secure AI agent platform using sandbox isolation for data protection. It supports one-click cloud deployment, integration with tools like Enterprise WeChat and DingTalk, and over 10,000 skills across models like Spark X2 and MiniMax-M2.5. Priced at 16.8 RMB/month with a promo offering unlimited dialogues until March 20.
💼
Hugging Face CEO Warns of AI Power Concentration
Hugging Face CEO Clement Delangue identified concentration of power as one of AI’s biggest risks, particularly if government intervention becomes excessive. He also discussed the recent OpenAI model hack involving Hugging Face and the challenge of keeping AI agents inside secure testing sandboxes.