📄 中文摘要
AI 编码智能体在生产环境中误删配置文件等事件，凸显了当前 AI 辅助开发中问责机制的缺失。现有版本控制系统仅记录代码变更，却无法追踪智能体决策的动因、指令来源以及事件链是否被篡改。这种问责鸿沟对受监管行业和企业安全构成重大风险，因为缺乏可信的审计日志，导致难以证明合规性、追溯安全事件或进行内部调查。AI 智能体的自主性日益增强，意味着其行为可能产生法律和财务后果，而传统聊天记录等证据易于伪造或丢失。因此，迫切需要一种防篡改的审计追踪方案，以确保 AI 智能体操作的透明度、可验证性和不可抵赖性，从而弥补 AI 辅助开发中的治理空白，保障企业在采用高自主性 AI 工具时的安全与合规。
📄 English Summary
InALign: Tamper-Proof Audit Trails for AI Agents
The increasing autonomy of AI coding agents, exemplified by tools like Claude Code, Cursor, and Copilot, introduces significant accountability challenges in software development. Incidents such as an AI agent deleting a production configuration file highlight a critical gap: current version control systems track code changes but fail to record the rationale behind an agent's decisions, the specific prompts it received, or whether the sequence of events has been tampered with post-factum. This accountability deficit is particularly problematic for regulated industries and enterprises, where verifiable audit trails are essential for demonstrating compliance, investigating security breaches, and conducting internal inquiries. Without tamper-proof logs, proving who instructed an agent, when, and why becomes impossible, leaving organizations vulnerable to legal and financial repercussions. The reliance on easily modifiable chat logs as the sole evidence further exacerbates this issue. A robust, immutable audit trail solution is imperative to ensure the transparency, verifiability, and non-repudiation of AI agent actions. Such a system would bridge the governance gap in AI-assisted development, providing the necessary security and compliance infrastructure for organizations embracing highly autonomous AI tools.