ClawHub Security: The Malicious Skills Attack Explained
ClawHub is the public registry where OpenClaw users find and install skills. In early February 2026, security researchers found hundreds of malicious skills on it. Most belonged to one campaign, named ClawHavoc, that installed a data stealer on macOS. OpenClaw responded within days with VirusTotal scanning and a reporting system. If you run OpenClaw or any agent that installs third-party skills, the same trick works on your setup. ClawHub security scanning still misses plain-language instructions, so you have to read skills yourself and watch your agent's memory files.
What is ClawHub, in one sentence?
ClawHub is the public registry for OpenClaw skills and plugins, where anyone with a qualifying GitHub account can publish and any OpenClaw user can install with one command.
The OpenClaw docs describe it as hosting versioned skills and plugins. Each entry shows a security scan summary next to its download count. A skill is mostly instructions. A SKILL.md file tells the agent what to do, sometimes with scripts alongside. So a harmful skill needs no code at all, just a sentence telling the agent to run something.
What happened in the ClawHavoc attack?
The Hacker News reported on February 2, 2026 that Koi Security had audited 2,857 ClawHub skills and found 341 malicious ones. Koi tied 335 of them to one campaign and named it ClawHavoc.
The trick was a fake prerequisites section. A skill posing as a crypto wallet tool or a YouTube utility told the user to install something first. On macOS that meant pasting a script into Terminal. The chain ended in a binary resembling Atomic Stealer (AMOS), a commodity macOS stealer. On Windows the lure was a password-protected archive on GitHub that held a keylogging trojan.
The count kept growing as others looked. Antiy CERT's analysis, with figures as of February 5, says at least 1,184 malicious skills had appeared on ClawHub. It attributes them to 12 author IDs, and one account uploaded 677. Antiy lists the AMOS targets as crypto wallets, SSH keys, browser passwords and Telegram sessions. Another sample read the agent's local .env file, which holds API keys for paid AI services.
How does a malicious skill differ from a malicious package?
Package registries like npm have dealt with typosquats and stealers for years. Skills add a new layer, because the payload can be plain English aimed at the agent.
| Malicious npm or PyPI package | Malicious agent skill | |
|---|---|---|
| Payload | Code that runs on install or import | Code, or instructions the agent follows |
| Who executes it | The package manager or your program | Your agent, with your agent's permissions |
| What scanners catch | Known malware signatures and suspicious code | Code payloads, rarely plain-language instructions |
| Persistence | Stays installed until removed | Can also write itself into the agent's memory and instruction files |
Persistence is the row with no package-registry equivalent. Snyk's ToxicSkills study, published February 5, 2026, warns that malicious skills can poison agent memory for persistence. Snyk advises reviewing SOUL.md and MEMORY.md for unauthorized changes. A skill you deleted can leave instructions behind in a file the agent reads every session.
How reliable are the numbers on malicious skills?
Every count in this story comes from a security vendor. Koi Security and Antiy are security firms, as is Snyk. Each used its own method and date. Of 3,984 skills Snyk scanned from ClawHub and skills.sh, 36.82% had at least one flaw and 13.4% had a critical one. Human review confirmed 76 malicious payloads, and 91% of those also used prompt injection. Across all of ClawHub, Snyk's prompt-injection detection rate was 2.6%.
Snyk's "any flaw" figure includes issues such as hardcoded keys. The strict figure is the 76 payloads that human reviewers confirmed. The jump from 341 to 1,184 reflects a later date and a different tally.
What did OpenClaw change after the attack?
On February 7, 2026, OpenClaw announced a VirusTotal partnership. Every published skill is bundled and hashed, then checked against VirusTotal. Unknown bundles get uploaded for analysis, including Code Insight, an LLM review that starts from SKILL.md. Benign skills are approved automatically. Suspicious ones carry a warning, and malicious ones are blocked from download. Active skills are re-scanned daily.
The post is candid about limits. It calls scanning "not a silver bullet" and says natural-language instructions to an agent may not trigger a virus signature. So a skill can pass the scan and still tell your agent to run an installer.
The Hacker News also reported a reporting feature from OpenClaw's creator, Peter Steinberger. Skills with more than three unique reports are hidden by default. The current docs say publishing needs a GitHub account old enough to pass an upload gate. Moderators can hide content and ban accounts, and scan-held releases can disappear from the public catalog.
How do you audit an OpenClaw skill before installing it?
Start with the scan result on the skill's ClawHub page, then go further than the scan:
- Read the whole SKILL.md. Any step that asks you or the agent to download and run something is a red flag, whatever the reason given.
- Check the publisher's account and other uploads. ClawHavoc used typosquats of ClawHub's own name.
- Look for code that reads .env files or posts to webhook sites.
- After installing, diff your agent's memory files against the last version you trusted.
- Keep API keys out of files the agent can read when the task does not need them.
The bottom line
Treat a third-party skill like a shell script from a stranger, because your agent will run it with your permissions. Use ClawHub's scans as a first filter. Then read SKILL.md yourself and prefer skills from publishers you know. Put your agent's memory files under version control so a poisoned instruction shows up as a diff.
How do you keep agent memory and skills auditable?
ClawHavoc showed two weak points in agent setups. One is skills written by strangers. The other is memory files that change without anyone reviewing them. Snyk's advice to check MEMORY.md only works if memory is a file you can read.
memU distills reusable skills from your own agents' history, as readable Markdown. A scheduled background task slices new session logs into jobs, and your agent writes the memory and skill files. memU's MemoryService makes no LLM or chat calls. It stores and embeds that Markdown and retrieves it when a standing instruction in CLAUDE.md or AGENTS.md asks. Hosts that memU supports, including OpenClaw, share one backend. Your agent writes the memory, so a poisoned session can teach it something bad. Because the memory is Markdown, you can search for that entry and delete it.
The hosted version is free and syncs across devices. A self-hosted install keeps memory private on one machine. For hosted, get an API key from memu.so, then send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is
YOUR_MEMU_API_KEY.
memU is open source under Apache-2.0, and the self-hosted steps are in the memU GitHub repository.
FAQ
Is ClawHub safe to use?
Safer than in January 2026. Every skill is now scanned with VirusTotal and re-scanned daily. OpenClaw itself says a clean scan is no guarantee.
How many malicious skills were found on ClawHub?
Koi Security found 341 in an audit of 2,857 skills, 335 of them from the ClawHavoc campaign. Antiy CERT later counted at least 1,184 that had appeared on ClawHub as of February 5, 2026.
What malware did ClawHavoc install?
On macOS, a binary resembling Atomic Stealer (AMOS), which goes after crypto wallets and saved credentials. Windows users were sent a password-protected archive containing a trojan.
Can a malicious skill change my agent's memory?
Yes. Snyk warns that malicious skills can poison agent memory to persist after removal. Review files like MEMORY.md and SOUL.md for changes you did not make.
How do I report a malicious ClawHub skill?
Signed-in users can report skills on ClawHub, and moderators review the reports. OpenClaw also takes security reports at security@openclaw.ai.