OpenClaw logs can’t prove most policy violations, new study warns

A new study posted to arXiv makes an uncomfortable point about autonomous AI agents. Most of their policy violations cannot be proven from logs alone.

The exception is narrow. Violations become provable only when approvals are recorded in structured, machine-readable formats. For OpenClaw, the fast-growing open-source agent framework, that is a problem. Its default logs were not built with that kind of evidence in mind.

What the logs actually capture

OpenClaw’s default logging works by writing tool calls to local JSONL files, a plain-text format where each line holds one record.

What those files leave out is the part compliance teams care about most. They do not record which policy rules were applied or the basis for a decision. They also skip the approval hierarchy behind an action.

The criticism is not limited to academic papers. Compliance-focused analyses have raised the same concern, as has GitHub issue #55801, opened March 27, 2026. That issue flagged that the default logs omit policy rule applications and decision foundations.

Reports from October 2026 pushed the argument further. They said activity logs for agentic systems must include intent, authority, and lineage data to prove policy compliance.

Those three terms are worth unpacking. Intent covers what the agent was trying to accomplish. Authority covers who or what allowed it. Lineage covers how one action led to the next.

How OpenClaw got here

OpenClaw started life as Warelay and was rebranded in late 2025. It built its following by letting autonomous agents handle tasks through messaging platforms and connected tools.

The framework has reached millions of instances, and it officially became a non-profit as the OpenClaw Foundation in July 2026.

Enterprises are asking whether it can meet standards like SOC 2, the audit framework many companies use to show customers their controls actually work.

The developers have not ignored the issue. A pull request, #85673, added bounded audit logs to the framework. The change supports better tool-policy filtering and documents when deny rules are matched.

Past Carnegie Mellon research has suggested that audit log completeness challenges predate modern AI agents. They have simply resurfaced as agents take on more autonomous work.

What this means for agent adopters

The most direct impact lands on organizations deploying OpenClaw in regulated settings. If logs cannot prove a violation, they also cannot convincingly prove compliance.

For OpenClaw specifically, the questions to watch are concrete. Will the bounded audit logs from pull request #85673 expand to capture approvals in structured form? Will the foundation publish guidance on meeting SOC 2 and similar standards? And will enterprise users treat the default setup as acceptable or as a starting point?