OpenClaw 2026.11.4: approval for delegated changes and fewer leaked secrets
What is in OpenClaw 2026.11.4? It is a patch release from November 16, 2026, with 125 pull requests from 90 contributors. The main change: when an agent wants to make persistent changes on someone else's behalf and that authority is restricted or unverified, OpenClaw now asks for approval first. Secrets also leak less often into logs, diagnostic output and summaries, and bugs that could drop messages or replies are fixed.
We read the release on GitHub and the full changelog. We covered the previous patch in our article on OpenClaw 2026.11.2.
What happened
The stable tag v2026.11.4 was published on November 16 at 04:17 UTC. The project summarizes it as fixes for three kinds of failure: buffered Signal messages skipped after a reconnect, replies lost after a settings reload, and damaged recovery copies of memory and configuration. It also adds a setting for approving delegated changes and lets policy plugins inspect images before they reach a model.
Details
| Area | According to the changelog |
|---|---|
| Delegated changes | Approval required when authority is restricted or unverified; new standing approval setting; a revoked autoApproveChanges grant is rechecked during a run |
| Secrets | No longer shown in Gateway and Doctor diagnostics, client logs, OpenTelemetry attributes or post-compaction summaries |
| Sandbox | PostgreSQL connection variables are no longer inherited by host and sandbox commands |
| Memory | Less memory pressure when indexing long transcripts; large SQLite rebuilds could fail |
| Scale | 125 pull requests, 106 direct commits, 90 contributors |
The security section is the longest. Examples: Doctor sometimes imported stale plaintext credentials although a stored reference existed, saving a skill key could replace an environment-backed key with plaintext, and screenshots from paired devices could be attached to outgoing messages automatically. Those cases are now closed. Commands that start with an environment assignment (NAME+=value) can no longer pass on allowlist approval alone.
What this means
For anyone running OpenClaw as a personal or team agent, the approval rule is the change you will notice: persistent changes an agent makes on someone else's behalf no longer slip through on an outdated automatic setting. It fits a wider move in agent software towards checking permissions per action; see our overview of agentic AI and the trust rules in Codex CLI 0.169.
One upgrade note from the changelog: if you use FS_SAFE_PYTHON_MODE or OPENCLAW_FS_SAFE_PYTHON_MODE without an explicit native policy, replace it with the matching NATIVE_MODE variable and keep the value. We described an earlier OpenClaw limit in the 8192-token output cap; more on agent risks is in our overview of AI and cybersecurity.
What we did not check
We did not install the release or read the pull requests ourselves. The commit count differs between the GitHub page (220) and the changelog (106 direct commits next to 125 pull requests); we use the project's own figures.
Checked on November 17, 2026.
This article was written with AI (Claude) and checked against the sources below.
Sources
- OpenClaw, release v2026.11.4 (November 16, 2026): github.com
- OpenClaw, changelog 2026.11.4: raw.githubusercontent.com
- OpenClaw, release notes 2026.11.4: docs.openclaw.ai


