卡内基梅隆论文：OpenClaw 日志无法证明 4 项策略违规中的任何一项
New Carnegie Mellon paper finds that OpenClaw's logs left 0 of 4 policy violations provable, a warning for any team relying on agent logs for audits.
卡内基梅隆大学新论文发现，OpenClaw 的日志无法证明 4 项策略违规中的任何一项，原因是审批与权限信息往往埋在聊天文本里。研究还发现只有 10% 到 30% 的违规运行需要人工判断，多数违规可用代码检查。作者在 3 个基准的 696 次违规运行上测试了可验证智能体保险方案，检查器证明了 42.2%；若采用结构化审批记录，作者估计可达 83.0%，每会话区块链费用低于 1 美分。
正文
当前显示原文，中文全文译文尚未就绪。
New Carnegie Mellon paper finds that OpenClaw's logs left 0 of 4 policy violations provable, a warning for any team relying on agent logs for audits.
They found that most agent policy violations in its tests could not be proven from logs, often because approvals and permissions were buried in chat text.
Also finds that only 10% to 30% of violating runs needed human judgment, so most agent rule breaks could be checked by code with better logs.
Safety work mostly tries to stop agents before they misbehave. When a mistake slips through, like a wiped inbox or a bad payment, users get little beyond slow, capped, case-by-case refunds.
In verifiable agent insurance, a provider posts a deposit on a blockchain and promises rules a program can check, like paying only approved merchants. A neutral verifier compares the agent's tool-call log with on-chain hashes, and if a rule was broken, the deposit pays the user.
Tested on 696 rule-breaking runs from 3 benchmarks, the checker proved 42.2%, often failing where approvals lived only in chat messages. With structured approval records, the authors estimate 83.0%, at under 1¢ of blockchain fees per session.
Design your agent so every permission, its scope, and each risky action land in a machine-readable log.