☁️AWS Machine Learning Blog•較早收集於 13m
使用 Chrome 企業政策控管 AI 代理瀏覽行為
💡透過在 Amazon Bedrock AgentCore 上強制執行企業瀏覽器政策，確保您的 AI 代理安全。
⚡ 30-Second TL;DR
有什麼變化
設定 Chrome 政策以限制代理僅能瀏覽特定網站
為什麼重要
為部署需要網路存取權的 AI 代理的企業提供了必要的安全防護機制。
下一步行動
將 Chrome 企業政策套用到您的 AgentCore 部署，以強制執行嚴格的網域白名單。
誰應關注：Enterprise & Security Teams
關鍵要點
- •設定 Chrome 政策以限制代理僅能瀏覽特定網站
- •使用工作階段錄製來稽核並觀察政策執行狀況
- •實作自訂根 CA 憑證以進行安全的企業瀏覽
🧠 深度解析
Web-grounded analysis with 24 cited sources.
🔑 增強重點摘要
- •The Amazon Bedrock AgentCore Browser operates within a secure, isolated containerized environment, ensuring web activity is separated from the main system and offering features like session isolation, live viewing, and CloudTrail logging for comprehensive observability.
- •Beyond Chrome enterprise policies, organizations can implement domain-based filtering for AgentCore using AWS Network Firewall, enabling precise control over which websites agents can access, blocking unwanted destinations, and logging connection attempts for audit and compliance purposes.
- •The AgentCore Browser provides advanced functionalities such as CAPTCHA reduction, support for custom browser extensions to automate complex workflows, session profiles, and proxy configurations, enhancing automation capabilities while maintaining security.
- •Fine-grained access control for AI agents is further enabled through deep integration with AWS Identity and Access Management (IAM) and AgentCore Identity, supporting delegated authentication flows and OAuth 2.0 for secure access to resources and third-party services.
📊 競品分析▸ Show
While direct feature-by-feature comparisons of AI agent web browsing control are not universally detailed across all platforms, several enterprise AI agent platforms offer robust governance and integration capabilities that are relevant for secure agent operations:
| Feature / Platform | AWS Bedrock AgentCore | Microsoft Copilot Studio | Google Vertex AI Agent Builder | IBM watsonx Orchestrate |
|---|---|---|---|---|
| Core Offering | Fully managed service for building and scaling generative AI applications with agents, including secure web browsing. | Low-code agent building platform within Microsoft 365 and Azure ecosystem. | Cloud-native platform for building and deploying ML and generative AI applications, including agents. | Enterprise AI agent platform for compliance-heavy sectors, combining FMs with data governance. |
| Web Access Control | Chrome enterprise policies, AWS Network Firewall for domain filtering (SNI inspection), containerized browser, custom root CAs, session isolation, logging. | Agents operate within Microsoft 365 tools (Teams, SharePoint, Outlook), leveraging Entra Agent ID for identity and access management. | Integrates with Google Cloud data stack; Chrome Enterprise Premium offers visibility into data transfers and blocking unsanctioned AI tools. | Focuses on governed AI for regulated workflows; model-agnostic architecture. |
| Security & Governance | IAM, AgentCore Identity (OAuth 2.0), Guardrails for content filtering, CloudTrail logging, session recording, zero operator access (ZOA) architecture for inference privacy. | Entra Agent ID for identity/access management, human-in-the-loop controls, depends on Microsoft 365 data organization. | Role-based access control, environment isolation, audit logs, data residency controls. | Strong data governance via watsonx.data and watsonx.governance, auditability. |
| Integration | Integrates with AWS services (S3, Lambda, VPC, CloudWatch, CloudTrail), supports open-source frameworks (Strands, LangGraph, AutoGen). | Power Platform connector library (1,400+ integrations), deep integration with Microsoft 365. | Tight integration with BigQuery and Google Cloud data stack. | Integrates with existing business systems and various AI models/automation tools. |
| Pricing Model | Consumption-based (inference calls, agent invocation not charged separately). | Credit model (e.g., 25,000 credits for $200). | Contact sales; often consumption-based. | Contact sales. |
🛠️ 技術深入
- Execution Environment: The Amazon Bedrock AgentCore Browser operates within a secure, isolated containerized environment. Each user session in AgentCore Runtime receives its own dedicated microVM, ensuring complete separation of compute, memory, and filesystem resources to prevent data leakage between sessions.
- Interaction Protocols: Agents interact with the browser using WebSocket-based streaming APIs. An "Automation endpoint" allows agents to perform actions like navigating websites, clicking elements, filling forms, and taking screenshots. A "Live View endpoint" enables real-time monitoring and direct user interaction with the browser session.
- Policy Enforcement Mechanism: Domain-based filtering for web access is implemented using AWS Network Firewall, which can perform SNI (Server Name Indication) inspection to enforce allowlists and blocklists, logging connection attempts for audit purposes. This provides a layer of defense-in-depth for network access control.
- Tool Integration & Frameworks: AgentCore provides managed tools, including the Browser and Code Interpreter. It is designed to be framework-flexible, supporting popular open-source agent frameworks such as Strands, LangGraph, Google Agent Development Kit (ADK), OpenAI Agents SDK, and Microsoft AutoGen.
- Customization and Extensibility: Users can choose between an AWS-managed browser or create a custom browser. The platform supports uploading Chrome-compatible browser extensions to Amazon S3, which are then automatically installed during browser sessions, enabling custom authentication flows, automated testing, and performance optimization.
- Security Features: Beyond network policies, the browser environment includes session isolation, built-in observability with CloudTrail logging, and session replay capabilities. Custom root CA certificates can be configured for secure enterprise browsing, and AgentCore Identity manages agent authentication and authorization using OAuth 2.0 flows.
🔮 前景展望AI analysis grounded in cited sources
Enhanced enterprise adoption of AI agents for sensitive tasks.
Robust security controls like restricted browsing, audit trails, and isolated environments address critical compliance and data privacy concerns, fostering trust for deploying AI in regulated industries.
Increased demand for specialized AI agent security solutions.
As AI agents gain more autonomy and web access, the need for sophisticated tools to manage, monitor, and secure their interactions, including detecting 'Shadow AI' and anomalous agent activity, will grow significantly.
Evolution of browser technology to natively support AI agent interactions.
The integration of AI agent controls into enterprise browsers like Chrome suggests a future where browsers are designed with AI agent security and functionality as a core consideration, potentially leading to new browser-level AI features and policies.
⏳ 時間線
2023-04
Amazon Bedrock announced.
2023-07
Agents for Amazon Bedrock introduced in preview.
2023-09
Amazon Bedrock generally available.
2023-11
Agents for Amazon Bedrock generally available with enhanced orchestration and visibility.
2024-2025
AWS expanded Bedrock to include AI agents, allowing models to interact with external systems.
2026-01
Amazon Bedrock AgentCore Browser supports custom Chrome browser extensions.
📎 來源 (24)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： AWS Machine Learning Blog ↗