The OpenClaw Security Crisis: Anatomy of the First AI Agent Meltdown
Three high-severity CVEs, 341 malicious skills in the marketplace and 21,639 instances reachable from the internet. What happened to OpenClaw in early 2026.
· Updated · 12 min read · Guard0 Team

In the last week of January 2026, OpenClaw went from a personal project to the most talked-about AI agent on the internet. It is a self-hosted assistant, created by Austrian developer Peter Steinberger, that connects a language model to your messaging apps and runs tasks on your machine. It launched as Clawdbot, was renamed Moltbot on 27 January 2026, and became OpenClaw a few days later.
Within the same few weeks, three high-severity vulnerabilities were patched in the gateway, a security audit found 341 malicious skills in its marketplace, and an internet scan counted 21,639 instances reachable from the public internet.
This post covers what is on the public record, with a source for each claim, and what it means for anyone running an agent platform with plugins.
What Is OpenClaw
OpenClaw runs on your own hardware. A gateway process handles orchestration, tool routing and memory, and by default it listens locally on TCP port 18789. The agent can read and send messages, call APIs, browse the web and run shell commands, depending on what you allow.
Its behaviour is extended through skills: packages anyone can publish to ClawHub, the project's marketplace. A skill is instructions plus optional scripts, and it runs with the agent's access.
That marketplace is where the largest part of the problem started.
The CVEs: Three High-Severity Vulnerabilities
Three vulnerabilities were fixed in OpenClaw version 2026.1.29. The Centre for Cybersecurity Belgium (CCB) issued a warning about the first of them.
CVE-2026-25253: one-click remote code execution through token exfiltration (CVSS 8.8)
The Control UI accepted a gateway URL from the query string and opened a WebSocket connection to it without asking, sending the stored authentication token along. A victim who clicked a crafted link handed their token to the attacker, who could then connect to the victim's gateway and run commands with the victim's privileges. It needed no prior access to the target, only one click.
CVE-2026-25157: command injection in the SSH handler (CVSS 7.7 from GitHub, 7.5 from NVD)
A flaw in how SSH connection details were handled allowed an attacker to inject operating system commands. The attack surface is narrower than CVE-2026-25253, because it depends on SSH remote features being used, but a successful attack runs commands with the privileges of the OpenClaw process.
CVE-2026-24763: command injection in the Docker sandbox through PATH (CVSS 8.8)
OpenClaw can run tool commands inside a Docker sandbox. Unsafe handling of the PATH environment variable when building those commands allowed command injection in the sandbox execution path. The sandbox is the control people rely on to contain a skill, so a flaw there matters more than its narrow trigger suggests.
Why the three matter together
The count is small. The coverage is the problem. One flaw is in authentication, one in operating system integration, one in sandboxing. An attacker needs only one of them against an unpatched instance.
More CVEs were assigned to OpenClaw in the weeks that followed. If you run it, check the project's security advisories and stay on a current version.
ClawHavoc: The Supply Chain Attack
While the gateway was being patched, researchers were looking at the marketplace.
Koi Security audited all 2,857 skills on ClawHub and found 341 malicious ones, as reported by The Hacker News. 335 of them belonged to one campaign, which the researchers named ClawHavoc.
How the skills masqueraded
The skills looked like things people wanted. The reporting lists typosquats of the ClawHub command line tool itself (clawhubb, clawwhub, cllawhub and others), crypto wallet trackers, Polymarket trading bots, YouTube summarizers, auto-updaters, and Google Workspace integrations.
Atomic Stealer (AMOS) deployment
The skill's documentation told the user to install a prerequisite first. On macOS, that meant pasting a script into Terminal, which fetched a second stage and then a binary consistent with Atomic Stealer, a commodity infostealer. On Windows, it meant downloading a password-protected archive containing a trojan with keylogging. The targets were exchange API keys, wallet private keys, SSH credentials and browser passwords.
The install step is the trick. People who use agent skills are used to setup instructions, so a "prerequisite" does not look unusual. A smaller number of skills hid reverse shell backdoors inside working code, and one exfiltrated the bot's credentials file to a webhook.
The detection gap
ClawHub was open by default. There was no automated security scanning of skills at the time of the audit. After the disclosure, Steinberger added a reporting feature: signed-in users can flag a skill, and a skill with more than three unique reports is hidden by default.
The number kept growing after the first report. Koi later updated its count to 824 malicious skills as the marketplace grew past 10,700 skills.
21,639 Publicly Exposed Instances
At the same time, internet scanners were counting OpenClaw gateways.
As of 31 January 2026, Censys identified 21,639 exposed instances by matching the HTML titles of the control interface under its earlier names. The count had been about 1,000 less than a week before.
The gateway listens on the local interface by default, so each of these was put on the internet by a configuration choice: a reverse proxy, a tunnel, a cloud instance with an open security group, or a bind address changed for convenience.
Censys noted that most of the instances it could reach still required a token to view and interact with them. That is where the CVEs come back in. A token is the only thing protecting those gateways, and CVE-2026-25253 is a way to steal a token with one click.
An attacker with access to a gateway can direct the agent to run any installed skill, read its memory, and use every integration it is connected to. On a version before 2026.1.29, the command injection flaws are available as well.
The 21,639 figure is a snapshot from one scanner on one day. Other scans in the following weeks, using different methods, reported higher numbers.
The Response
The three CVEs were fixed in version 2026.1.29. The malicious skills Koi found were reported to ClawHub for removal, and the reporting feature described above was added to the marketplace.
Patching fixes the gateway. It does not undo what an installed skill already took. Credentials, API keys and wallet files stolen by an infostealer have to be treated as compromised and rotated, and a machine that ran a ClawHavoc skill needs a proper investigation.
Lessons for the Industry
None of this is unique to OpenClaw. Every lesson applies to any agent platform that runs third-party code.
Agent marketplaces need security review before publishing.
App stores work, imperfectly, because someone reviews what gets published. That raises the cost of publishing malware. ClawHub required a week-old GitHub account, and 341 of 2,857 skills turned out to be malicious.
Any platform that lets third-party code run inside an agent needs static analysis, sandboxed behavioural testing, publisher reputation, and a fast takedown process. This is baseline security for a software distribution system.
An exposed agent is worse than an exposed web app.
When an exposed web server is compromised, the attacker gets the web application. When an exposed agent is compromised, the attacker gets every tool the agent is connected to, everything it remembers, and the ability to act as the agent in every connected system.
A local-only default is the right default. The exposure happened one deployment at a time, when people moved the gateway onto a network without anything telling them what that meant.
Agent platforms need security testing before they are popular.
OpenClaw's adoption arrived in about a week. The security scrutiny arrived in the same week. Platforms at this layer handle credentials, hold sensitive memory and take actions for users. They need threat modelling and penetration testing before a major release, because fixing a widely deployed agent platform under public pressure is much harder.
The supply chain of skills, tools and integrations is the attack surface.
The most damaging attacks came through the marketplace: third-party code running inside the agent's trust boundary. Every plugin and every skill is a possible supply chain attack. What matters is whether you can see what is installed and tell when it is malicious.
For more on supply chain risk in the agent ecosystem, see AI Agent Supply Chain Security.
Scanning for OpenClaw Exposure with g0
g0, Guard0's open-source scanner, covers OpenClaw in three ways.
npx @guard0/g0 checkg0 check looks at the AI tools, MCP servers and installed skills on a machine and compares them against a known-malicious database that includes ClawHavoc campaign indicators.
npx @guard0/g0 scan ./my-openclaw-agentg0 scan reads OpenClaw's own files as part of a normal scan:
SKILL.md: prompt injection, permission escalation, and known ClawHavoc indicatorsSOUL.md: identity replacement and hidden directivesMEMORY.md: stored credentials, personal data, and injected trust overridesopenclaw.json: unofficial registries, hardcoded credentials, and unsafe runtime settings
g0 test --attacks openclaw-attacks runs OpenClaw-specific adversarial payloads against a live agent, including probes for CVE-2026-25253.
Scans run locally and produce a report with a severity and a fix for each finding. g0 does not upload your scans.
For the rest of what g0 does, see Introducing g0 or the g0 page. For how Guard0 finds agents like OpenClaw across a fleet of laptops, see Workforce AI.
Live walkthrough of agent discovery, risk scoring, and policy enforcement.
Book a Demo- Three vulnerabilities fixed in OpenClaw 2026.1.29: CVE-2026-25253 (CVSS 8.8, one-click RCE through token exfiltration), CVE-2026-25157 (SSH handler command injection), CVE-2026-24763 (CVSS 8.8, Docker sandbox command injection)
- ClawHavoc: 341 malicious skills out of 2,857 audited on ClawHub, 335 of them deploying Atomic Stealer through fake prerequisite installs. The count later rose to 824
- 21,639 OpenClaw instances reachable from the internet on 31 January 2026, most protected only by a token
- The causes were structural: an open marketplace with no security review, and gateways moved onto public networks
- The same lessons apply to every agent platform with plugins or a skill marketplace
Scan Your Agents
If you run AI agents, whether OpenClaw, another framework, or your own, start by finding out what is installed.
Run g0 on a machine or a repo:
See every agent in your company, what it can reach, and who owns it:
Many of the people who exposed a gateway or installed a bad skill were competent. They had nothing that told them there was a problem.
OpenClaw · Threat Intelligence · Vulnerability · AI Agents · Security Crisis