用Rust重写OpenClaw，Transformer作者下场造了安全版「龙虾」

Core Viewpoint - The article discusses the development of IronClaw, a secure version of OpenClaw, addressing significant risks associated with user data and funds exploitation [1][2]. Group 1: Background and Motivation - Illia Polosukhin expresses excitement about the emergence of OpenClaw, viewing it as a long-awaited technology that transforms human-computer interaction [4]. - Concerns are raised regarding the security of OpenClaw, highlighting risks such as credential leaks, prompt injection attacks, and potential fund theft [5][6]. - The decision to create IronClaw stems from a desire to prioritize user privacy and security, as the current convenience offered by AI systems may compromise personal and family safety [6]. Group 2: IronClaw Features - IronClaw is an open-source runtime environment designed with security as a core principle, built using Rust for clear and auditable code suitable for enterprise use [9]. - Key components of IronClaw include: - Agent Loop: Main message handling and job coordination - Router: Classifies user intent - Scheduler: Manages parallel job execution - Worker: Executes jobs with reasoning and tool calls - Orchestrator: Manages container lifecycle and authorization - Web Gateway: User interface for interaction - Routines Engine: Automates background tasks - Workspace: Persistent memory and retrieval - Safety Layer: Defense against prompt injection and content sanitization [10]. Group 3: Security Enhancements - IronClaw implements several critical changes to ensure security, such as: - Transitioning from direct file system access to database usage with controlled data policies - Utilizing WASM for dynamic tool loading in a sandbox environment - Preventing credential leaks through encrypted storage and strict access policies - Defending against prompt injection attacks with heuristic methods and future model updates - Employing a memory system with hybrid search to avoid file system damage [12][13]. - Future plans for IronClaw include regular updates, support for various communication channels, and the introduction of behavior policies for agents to ensure expected actions [13].