← Back to OpenClaw News An editorial illustration of a walled courtyard where agent templates arrive as glowing building kits and pass an approval archway, beside two freshly tagged release stones
November 19, 2026Release WatchSecurityEnterpriseSkills

OpenClaw Flags a High-Severity Approval Bypass as 2026.11.6 and a 2026.11.10 Beta With Agent Templates Get Tagged

A new high-severity advisory says restricted agents could skip approval through delegation in older builds. OpenClaw also signed a 2026.11.6 stable tag and a 2026.11.10 beta that adds agent templates from ClawHub, though neither is on npm yet. Main adds selective MCP approval and Codex certificate pinning, and Microsoft maps OpenClaw inside Autopilot.

OpenClaw published a high-severity security advisory on Wednesday evening: in older builds, a restricted agent could use delegation to change configuration without approval. Before dawn on Thursday the project also pushed two signed release tags to GitHub, a 2026.11.6 stable candidate and the first beta of a 2026.11.10 line that installs whole agents from ClawHub. Neither tag had a GitHub release page or an npm build when we checked at about 6:25 AM, so treat both as announced but not yet installable. Overnight, main let operators keep approved MCP tools on the worker, pinned remote Codex app servers to a certificate, and fixed pnpm updates that left machines with no openclaw command. Separately, Microsoft's Omar Shahine published a diagram showing where OpenClaw sits inside Autopilot. All times below are Pacific.

Advisory: Delegation Could Skip Approval for Restricted Agents

The OpenClaw security advisory GHSA-pw3x-g8v8-fvj2 on GitHub was published at 6:01 PM Wednesday and rated high, with a CVSS 3.1 score of 8.8. It says that in affected versions "the openclaw tool could apply persistent configuration changes without approval when tool policy denied execution and file writes." The impact section warns this "could let a restricted agent enable host command execution through configuration changes," with practical risk depending on configuration and on whether lower-trust input can reach that path. The listed affected ranges are 2026.9.2 up to but not including 2026.10.33, and 2026.11.1 up to but not including 2026.11.4. The first patched stable is 2026.11.4, and extended-stable users can move to 2026.10.33. The advisory does not list a CVE ID.

2026.11.6 Is Tagged, With Upgrade and Credential Fixes

The signed v2026.11.6 tag landed at 5:50 AM. The 2026.11.6 changelog file at the v2026.11.6 tag leads with what it calls smoother upgrades. It bridges the 2026.11.5 updater to the current release path and validates staged plugin dependencies correctly on Windows. Other headline items keep subagent and background turns running through configuration edits, and deliver final answers when a Gateway restart interrupted sending. The notes also promise lower startup memory pressure on constrained hosts and a fix that keeps plugin tool discovery from blocking the Gateway.

The fixes list has two credential items. Configuration reads now redact inline environment credentials, and SecretRef-backed skill API keys stay as references instead of being replaced with plaintext. PostgreSQL users get a guard that stops an upgrade before it replaces a still-serving installation when the PostgreSQL backup cannot run. Large file reads also stop scanning after the requested page instead of loading the whole file. The audited contribution record counts 189 pull requests in the range since 2026.11.5. At 6:25 AM, npm view openclaw dist-tags still showed latest and beta at 2026.11.5 and extended-stable at 2026.10.33. The docs page for 2026.11.6 release notes returned a 404.

SEN-X Take

The credential fixes are the reason to schedule 2026.11.6 once it reaches npm. A config read that echoes inline keys, or a skill key quietly turned from a SecretRef into plaintext, is exactly the leak an agent can repeat into a transcript. Until the build is published, grep your config for raw keys yourself. Wait for the npm tag instead of building from the tag.

2026.11.10 Beta 1 Brings Agent Templates From ClawHub

The v2026.11.10-beta.1 tag arrived earlier, at 4:22 AM. The 2026.11.10-beta.1 changelog at the beta tag puts agent templates first: add agents "directly from ClawHub packages, with a single CLAW.md manifest and bundled skill folders checked by the native trust flow." Its other new changes include a guest member role with sandboxed session tools, an opt-in usage-ranked tool index that covers MCP tools, a Settings choice for the Talk speech-to-text model, and stage and byte progress for long backups. Several of its highlights, such as browser handoff and update rehearsal, reach a beta tag after we covered them on main earlier this month.

The beta also comes with an unusual release-process note. Pull request #201802, which fast-tracks 2026.11.10-beta.1 qualification, merged at 4:17 AM at a maintainer's request. It records qualification for this one version and candidate as "waived, not passed," and states that "Failed/cancelled/unfinished coverage remains a known release risk." It also says "Package integrity and publication authority are not waived," and that the PR "does not itself publish the release."

For a look at the headline feature, the ClawCast episode list on openclaw.ai shows Episode 17, "Exploring Agent Templates in OpenClaw," dated November 18. In it, hosts Hannes Rudolph and Patrick Erichsen and guest maintainer Dan Riccio "try agent templates in Labs and talk about making OpenClaw easier to start with."

SEN-X Take

Agent templates change what a ClawHub install means. Until now you added one skill at a time, but a template brings a manifest plus a folder of skills in a single step. That makes the trust flow the real control point. Because this beta's qualification was openly waived, we would test it on a spare Gateway, not on an agent that handles customer mail.

Main: Selective MCP Approval, Swarm Progress and More Cron Room

Pull request #201081, which keeps allowed MCP tools on the worker, merged at 3:23 PM Wednesday. Before it, registering a Gateway policy for MCP tool calls routed every tool through the Gateway even when only a few needed human approval. A new optional allowLocalMcpTool callback can admit an exact node, server and upstream tool for same-worker execution, and "only a literal true permits local placement." Narrowing a policy revokes existing local grants, and existing policies keep their behavior unless they opt in.

Two more changes help people running many agents. Pull request #201382, which shows live Swarm progress inside chat turns, adds a compact grid of queued, running and finished workers to the turn that launched a Swarm. Selecting a worker opens it in Subagents. Pull request #197394, which adds up to eight cron runs past the default when memory is free, keeps the old cap of eight concurrent automation jobs. When all eight are busy, it starts one more per queue drain if at least 1 GiB of memory is free, up to 16. It reads memory with Node's process.availableMemory(), which respects container limits, and adds no new setting.

Overnight, pull request #201757, which keeps scheduled work with cron across Gateway restarts, stopped restarts from telling a conversation owner that a request they never sent "could not resume." On the model side, pull request #200468, which adds Google's nano banana 2.1 image model, makes google/gemini-nano-banana-2.1 the default for Google image generation. It supports up to 14 reference images, panoramic ratios such as 8:1, and 1K, 2K and 4K output.

pnpm Users: Read This Before Your Next Update

Pull request #201735, which keeps the openclaw command working after updating a pnpm install, merged at 4:24 AM. It describes real Docker tests in which updating a pnpm add -g openclaw install from 2026.11.4 to 2026.11.5 left every openclaw command failing with Cannot find module. On some pnpm versions the Gateway also refused to restart after its databases had been migrated. The fix restores the previous launchers instead of reporting success when the new package cannot start. The PR notes that the installed updater runs the relocation step, so 2026.11.4 and 2026.11.5 still publish a broken launcher when they update. The new package's Doctor and Gateway startup are designed to repair that launcher afterward.

A sibling fix, pull request #201701, which stops openclaw update from bricking a pnpm-linked Git checkout, covers developers who link openclaw to a source checkout and then ask for --channel beta or stable. On 2026.11.5 that switch can leave the Gateway down. With the fix, the updater refuses before it stops or changes anything. The PR is explicit that a checkout still on 2026.11.5 "can still start the switch and end bricked."

Security Tip: Close the Delegation Gap, Then Pin Remote Codex Servers

  • Check your version against the advisory. Run openclaw --version. If it falls in either range listed in GHSA-pw3x-g8v8-fvj2, move to 2026.11.4 or later on stable, or 2026.10.33 on extended-stable. Before you can upgrade, the advisory says to set tools.exec.mode to ask and deny untrusted delegated changes. It also recommends narrow channel and tool allowlists and not sharing one Gateway between mutually untrusted users.
  • Pin the server, not just the address. Pull request #201499, which verifies the admitted Codex app server before sending credentials, merged at 4:07 PM Wednesday. A Kubernetes Service can point to a different Pod after the original app server exits. Operators can now put the admitted server's PEM certificate in appServer.tlsPeerCertificate, and OpenClaw "rejects a different listener before sending credentials." In the PR's test, a replacement Pod at the same address received zero HTTP requests.
  • DefenseClaw users should be on 1.0.9. Cisco's DefenseClaw 1.0.9 release notes on GitHub, published November 12, say "OpenClaw no longer approves its own scopes or sends tokens to an unproven listener." The notes add that when the DefenseClaw gateway is down, default and permissive tiers allow calls with a warning while strict and managed tiers block them, so pick the tier on purpose. Setup is covered in the DefenseClaw OpenClaw connector guide.

Skill Card: Hashference Provider Setup for OpenClaw

The Hashference for OpenClaw skill page on ClawHub, published by Matthew Poletiek, describes a skill to "set up, use, refresh, or troubleshoot Hashference as an OpenClaw inference provider." When we checked, it listed version v1.0.1, an MIT-0 license and a ClawHub security audit result of "Pass."

  • What stands out. Its catalog helper makes a public HTTPS GET to the Hashference catalog and, per the page, "reads no environment variables, credentials, account files or OpenClaw configuration, writes no files, executes no child processes." Authentication goes through an existing SecretRef or OpenClaw masked entry, and the skill says a token must never go into "a model-visible response, log, transcript or source file."
  • Spend guardrails. The skill says purchases, staking, mining and paid inference need the user's authorization, and that "device approval is not spend authorization." It merges model entries without replacing other providers and leaves your primary model unchanged unless you ask.
  • Our caveats. Hashference is a paid inference service, and we have not tested it or its pricing. We read the skill page; we did not install it. Skill authors should also note pull request #201692, which stops the skill validator from rejecting skills that declare a version. Before it, a SKILL.md built from ClawHub's own version: 1.0.0 example failed to package.

Microsoft Shows Where OpenClaw Sits in Autopilot

On Tuesday at 9:19 PM, Omar Shahine posted Omar Shahine's Autopilot architecture diagram on X with the line "This is our high level architecture for Microsoft Autopilot and where @openclaw fits in." The diagram places an "OpenClaw Gateway stable image" inside a container instance labeled "untrusted sandbox," alongside persistence for workspace, state and memory. A "trusted multitenant service" called the Autopilot Gateway sits in front, with a proxy box labeled "Inspection, auditing, DLP, credential injection." Durable storage is drawn as the agent's and the manager's OneDrive.

That matches the sandbox defaults OpenClaw documents for self-hosters. The OpenClaw Docker sandbox backend documentation lists network: "none", readOnlyRoot: true and capDrop: ["ALL"] as defaults.

SEN-X Take

The useful lesson in Shahine's diagram is the label, not the logo. Microsoft treats the OpenClaw runtime as untrusted and keeps credentials on the trusted side of a proxy, injected per call. Teams self-hosting can copy that shape today: sandboxed sessions, no raw keys inside the container, and an inspecting proxy in front of outbound calls.

Community and What We Left Out

X posts on Thursday morning promoted Cisco's Skill Scanner and DefenseClaw alongside a claim that 20% of one skill marketplace was malware. We could not find that figure in Cisco's DefenseClaw repository, docs or 1.0.9 notes, so we did not use it. An X recap of NVIDIA SkillEvaluator 1.0 repeated what we covered yesterday. The 2026.10.33 extended-stable release and the Autopilot rollout were also covered yesterday. We skipped a report that GitHub briefly hid the Orca orchestrator's organization, since it is not an OpenClaw project, along with community posts that were jokes or anecdotes we could not check.

What to Watch Next

Watch for any follow-up to the delegation advisory, GitHub release pages and npm tags for 2026.11.6 and 2026.11.10-beta.1, the docs.openclaw.ai notes for both, and whether the beta's waived qualification reruns cleanly before a second beta.

Patching the Approval Bypass or Testing Agent Templates?

SEN-X helps teams audit delegated-agent approvals, stage OpenClaw betas on spare Gateways, pin remote Codex app servers, and design MCP approval policies that keep routine tools fast.

Contact SEN-X →