← Back to OpenClaw News An abstract editorial illustration of a glowing software release package rising on a beam of light above a library of skill cards, one card held under a magnifying lens beside a fading campaign map
November 15, 2026Release WatchSecurityBrowserPlugins

OpenClaw 2026.11.5 Beta 3 Lands as a Skill Named in February's Stealer Campaign Stays Listed on ClawHub

OpenClaw 2026.11.5-beta.3 is on npm with fixes for long sessions, restarts and compaction. A skill named in Trend Micro's February stealer report is still listed on ClawHub with a clean verdict, main adds experimental Kitesurf browsing, and automations get their cross-provider fallbacks back.

OpenClaw's weekend brought a new beta and an awkward security finding. Version 2026.11.5-beta.3 reached npm's beta channel on Saturday afternoon, focused on long sessions, restarts and recovery. It also adds a setting that lets compaction summaries grow beyond 16,000 characters. On the same day, an independent research site said a ClawHub skill named in a February malware report is still listed with a clean moderation verdict. We checked what we could against ClawHub's own public data. Main also gained an experimental Kitesurf browser engine and several automation fixes, a ClawHub plugin brought TensorFold local models to OpenClaw, and the project account said plainly on X that OpenClaw is "not a business." All times below are Pacific.

2026.11.5 Beta 3 Reaches npm With Long-Session and Restart Fixes

The OpenClaw 2026.11.5-beta.3 release notes on GitHub were published at 4:31 PM on Saturday, and npm now lists 2026.11.5-beta.3 under the beta tag. The latest tag is still 2026.11.3 and extended-stable is still 2026.9.37, so nobody on the default channel gets this build without opting in. The audited contribution record covers 50 pull requests since beta 2.

The notes group the work under three highlights. Large-session reliability reduces memory pressure during transcript indexing, trimming and compaction. One of the listed fixes, pull request #194099, which avoids worker out-of-memory crashes during large transcript index repair, explains the failure: the repair path built an array of every decoded search text before any byte budget applied, so a large transcript could end its 512 MiB worker with ERR_WORKER_OUT_OF_MEMORY. It now rereads texts lazily in batches. The PR is careful to say it does not prove that every ordinary append incident is resolved.

Restart and delivery safety is the second theme. Pull request #196540 preserves queued send permissions across Gateway restarts. Before it, a queued send could resume under the sender's broader current permissions after a restart, instead of the limits of the source that first accepted it. Each queued send now keeps its original app, membership, scope and model limits, and a missing or revoked source turns into a failed send that can be resubmitted. The third highlight, recovery and setup, covers cold-start initialization, interrupted workspace recovery, installer permissions and clearer iMessage Automation status.

The change most operators can act on is a new compaction ceiling. The OpenClaw compaction documentation now says client-side compaction keeps at most 16,000 characters of summary by default, and that agents.defaults.compaction.maxSummaryChars or a per-agent override accepts integers from 1,000 to 200,000. The docs add a warning worth repeating: "This is a ceiling, not a target." Larger summaries raise context use and replay cost, and they do not recover facts already lost from earlier summaries. The notes also list Lightpanda navigation through its supported connection path, Slack reasoning-stream freeze fixes, Windows case-insensitive state paths and macOS LaunchAgent recovery when Trash access is denied.

SEN-X Take

Beta 3 reads like a stabilization build rather than a feature drop, which is what you want this close to a stable 2026.11.5. The queued-send fix is the one we would test first: it closes a real gap where a restart could quietly give old work more authority than it was accepted with. Raise maxSummaryChars only for agents that truly lose context, and watch token costs when you do.

A Skill Named in February's Stealer Campaign Is Still Listed on ClawHub

On Saturday, Skillcatraz, an independent site that says it detonates suspicious skills and watches what they do, published a Skillcatraz finding on the speckit-coding-agent ClawHub skill and flagged it to the project on X. The site says the skill was attributed by security vendors in February 2026 to an Atomic macOS Stealer (AMOS) campaign that used a fake command-line prerequisite. It reports that 39 of the 40 skills named in those February reports have since been removed, and this is the one still listed. In its sandbox test on November 14, Skillcatraz says an agent following the skill's setup text reached the campaign's published staging host, which now answers HTTP 451, so no payload was captured. Skillcatraz calls it "a takedown-grade registry finding, not a new-malware capture."

We checked the parts that can be checked without running the skill. The Trend Micro research report on malicious OpenClaw skills used to distribute Atomic macOS Stealer, dated February 23, 2026, describes malicious instructions hidden in SKILL.md files that present fake setup requirements. Its Trend Micro indicator list of 40 OpenClaw skills names heldinhow/speckit-coding-agent first. On Sunday morning we queried ClawHub's public skills API for all 40 names. Thirty-nine returned not found. The speckit-coding-agent entry returned a live listing: owner heldinhow, latest version 1.1.2 (published February 13), moderation verdict clean and isSuspicious: false. Skillcatraz also says the skill's latest security scan now reads suspicious; the API fields we read do not include that scan, so we could not confirm it.

By about 6:30 AM Sunday we had found no public response from OpenClaw or ClawHub on X, and no related issue in OpenClaw's GitHub repositories. We did not install, download or run the skill. On November 10 we covered a separate Acronis report on ClawHub malware campaigns; this is a narrower and more uncomfortable point, because one named indicator appears to have slipped through cleanup for nine months.

SEN-X Take

A clean verdict on a skill that sits on a published vendor indicator list is a process gap, not a scanner bug. Instruction-only lures carry no code to flag; the harm is one fetch away. Registries should cross-check public IOC lists against live listings. Until ClawHub acts, teams should block this slug and treat "clean" as one signal among several, not a clearance.

Kitesurf Joins the Browser Lineup as an Experimental Engine

At 6:47 AM Saturday, main merged pull request #198833, which adds experimental Kitesurf browsing to OpenClaw's browser profiles. Kitesurf is Cloudflare's newly open-sourced browser engine. The updated OpenClaw lightweight browsers documentation says it parses HTML, applies CSS, runs JavaScript and renders images in Workers without a Chromium process, and that its CDP endpoint implements only a subset of Chrome's protocol. Some unsupported commands return empty success responses, which is why OpenClaw does not treat it as Chrome.

You opt in with a profile that sets engine: "kitesurf", a cdpUrl and attachOnly: true. OpenClaw never installs, launches or deploys the engine. The first adapter admits navigation, tabs, text, AI snapshots, semantic click, type and fill actions, and experimental screenshots. PDF, downloads, uploads, dialogs, emulation and action batches stay off. The docs say not to use this profile to keep logins, note that hosting on Cloudflare needs the Workers Paid plan, and warn that upstream has no built-in authentication, so a deployed endpoint must be protected.

Automations Get Their Fallbacks Back and Hear About Failures Sooner

Three changes on main touch scheduled work. Pull request #196539 lets configured fallbacks on other providers run for account-bound automations. It merged on Thursday, November 12, and is not listed in beta 3's notes. The PR says that since an earlier change, a scheduled job recorded one inference account and refused every model on any other provider, so a quota failure, outage or rate limit on the primary provider failed the job. Cross-provider fallbacks now run again with no Doctor step, including for jobs stored by 2026.10.10 and 2026.11.2. The job's recorded account still governs its own provider, which the PR says keeps the original billing guarantee.

Pull request #197704 lets each scheduled job keep its own Access selection (11:13 AM Saturday). Jobs can keep Standard, Daybreak Blue or Daybreak Red through openclaw automations add --access or edit --access, without changing the conversation's setting. The PR stresses that "the saved choice grants no permission"; roles, accounts and runtimes still decide. Early Sunday at 3:43 AM, pull request #199558 sends an automation's first failure to its owner agent. Its example is a weekly job whose check script hit a GitHub 403: repair waited on the default alert threshold of two failures in a row, so the agent would hear about it after a week and the user after two.

SEN-X Take

If your overnight jobs failed during a provider outage this month despite configured fallbacks, #196539 is the likely explanation, and it is not in a tagged release yet. Check run history for refusals that name an account, and keep a second provider configured. The first-failure change matters most for weekly jobs, where a two-strike alert rule meant two lost weeks.

Also on Main

From the Maintainers on X

Replying late Saturday to a long-time user who said OpenClaw "feels like it's falling behind" and asked for free models, the official OpenClaw account's reply on X asked what specifically felt behind and said: "We don't sell tokens or give out credits or free models. We're not a business." Earlier that afternoon, in a thread about harnesses, Peter Steinberger's post on X said, "We build OC with OC now; that move happened a few months ago." Both are short replies, not announcements, but they state the project's model and its dogfooding plainly.

Security Tip: Read a Skill's Setup Steps Before Your Agent Does

The speckit-coding-agent case shows why a registry verdict is not enough. The Trend Micro AMOS report describes skills that hide instructions in SKILL.md and exploit AI agents as "trusted intermediaries" that present fake setup requirements, followed by a deceptive dialog box that asks for your password. Skillcatraz notes this skill ships no code at all, so the harm sits in its setup text.

  • Read prerequisites line by line. Before installing any skill, open its SKILL.md and look for setup steps that fetch and run something from a host the skill names. Trend Micro's indicator list notes that base64 -D | bash is a sign of a possibly malicious skill.
  • Never type your password for a skill's setup. A login or keychain prompt that appears while an agent follows a skill's instructions is the exact pattern Trend Micro describes.
  • Check published indicators, not just verdicts. Compare skills you use against vendor IOC lists, and remove heldinhow/speckit-coding-agent if it is installed anywhere you manage.

Plugin Card: TensorFold Provider for OpenClaw

For teams that want agents on local hardware, the TensorFold provider plugin listing on ClawHub (v0.2.2, Apache-2.0, security audit marked Pass) makes TensorFold an OpenClaw model provider. Its README says TensorFold serves language models from a single Zig binary on Apple Silicon and NVIDIA GPUs, with no Python or MLX to install. Install it with openclaw plugins install clawhub:openclaw-tensorfold, then run openclaw onboard --auth-choice tensorfold-local.

  • Starts on demand. OpenClaw launches tensorfold serve on 127.0.0.1 for the first request that needs it, and stops it when OpenClaw exits or after an idle time you choose.
  • Real model limits. The provider reads /v1/models for each model's context window, image support and reasoning levels instead of guessing from the name. Setup defaults to a 65,536-token window.
  • Private by default. The server and OpenClaw share one key in tensorfold/server.keys (mode 0600). The README lists checks on an Apple M5 Max with OpenClaw 2026.11.3, and says the idle stop inside a long-running Gateway and a CUDA host are not checked yet. It needs OpenClaw 2026.11.3 or later.

What We Checked and Left Out

An X post said issue #177506 had an official fix, but GitHub still shows the issue open, so we left it out. A commit shared on X turned out to be a November 2 change we did not need to revisit. We did not use a third-party AgenticBench score for OpenClaw 2026.11.3 without reviewing its method, or a Japanese security roundup that was secondary to the release notes we read directly. No GitHub security advisories were published for OpenClaw in the window, and 2026.11.4 has not shipped.

What to Watch Next

First, whether ClawHub removes speckit-coding-agent and explains how a listed indicator kept a clean verdict. Second, whether 2026.11.5 reaches stable with the beta 3 fixes, and whether the cross-provider fallback fix rides along. Third, how Kitesurf holds up once people test it against real sites, and whether OpenClaw widens what the adapter allows.

Vetting the skills and plugins your agents install?

SEN-X helps teams review ClawHub skills against published indicators, set up safe automation fallbacks, and test OpenClaw betas on staging before they reach production agents.

Contact SEN-X →