Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore

Odin Research / What is / AI tools and agents

What Is OpenClaw? The Self-Hosted AI Agent and Its Security Risks

A hugely popular personal AI agent that can run tools, drive a browser and talk through chat apps. Here is how it works, where it goes wrong and what Odin Research found exposed online.

By
Odin by Brandefense
Updated
21 Nov 2026
Reading time
8 min read
Type
AI agent

In short

OpenClaw is an open-source, self-hosted AI assistant that runs on your own computer and takes actions for you through chat apps such as WhatsApp, Telegram, Slack and Discord. A local Gateway manages its sessions, tools and channels. Its power to run tools and third-party skills makes it a target for key theft and supply chain abuse.

Key takeaways

What it is
An MIT-licensed personal AI agent, earlier known as Clawdbot and Moltbot, that connects a language model to your files, shell, browser and more than 20 chat channels.
How it works
A local Gateway is the control plane for sessions, tools, events and channels. Skills from the ClawHub marketplace extend what the agent can do.
Why it matters
Unit 42 documented malicious ClawHub skills that delivered macOS infostealers, and NVD lists CVE-2026-25253, a flaw that sent a gateway token to an attacker-chosen address.
What Odin Research saw
932 secret and session matches in exposed OpenClaw folders on 21 servers, and a workbench that pointed OpenClaw at a logged-in Chromium profile through headless Playwright.
What to do first
Block web access to hidden agent folders, keep the Gateway private, update past 2026.1.29, vet every skill and rotate keys that ever sat in agent logs.

What is OpenClaw?

OpenClaw is an open-source AI assistant that runs on your own computer and acts for you through the chat apps you already use, such as WhatsApp, Telegram, Slack, Teams, iMessage and Discord. A local Gateway connects a language model to tools, files and channels. That reach is what makes it useful and risky.

The project's tagline is "The AI that really does things." It is MIT-licensed, created by Peter Steinberger and the community, and stewarded by the OpenClaw Foundation. NVD records that OpenClaw was also known as Clawdbot and Moltbot.

OpenClaw is one of the most popular AI agents on GitHub. The repository showed about 395,000 stars and 82,900 forks on 21 November 2026. It runs on macOS, Linux, Windows and through native mobile apps, and it can even run as a shared multi-user agent.

How OpenClaw works

OpenClaw has three main parts:

  1. The Gateway. The project calls it the local control plane for sessions, tools, events and channel connections. The Control UI, command line and terminal interface all connect to it.
  2. Channels. More than 20 messaging apps can carry instructions to the agent and answers back to the user.
  3. Skills and plugins. These add abilities, from browsing to trading. Skills are listed on ClawHub, the project's marketplace. Unit 42 describes them as markdown-driven packages with broad local system access.

The project ships an install script and an npm package, and it requires a current Node.js release. Its README sets a few security defaults. It tells users to treat inbound messages as untrusted input. Unknown senders on direct-message channels must be paired and approved. Tools run on the host for the main session unless sandboxing is configured.

On disk, OpenClaw keeps its state in a hidden .openclaw folder. Odin Research found that this folder holds configuration, provider keys in auth profiles, session logs and step-by-step trajectory files. Those files matter most when the folder ends up on a web server.

How attackers abuse OpenClaw

OpenClaw is not malware. Attackers target it, and some use it as a tool.

  • Malicious skills. In June 2026, Palo Alto Networks Unit 42 reviewed ClawHub abuse. It cited Koi Security's ClawHavoc disclosure of 341 malicious skills and a Bitdefender estimate that about 17% of skills it analyzed carried malicious payloads. Unit 42 found five more skills that passed ClawHub's screening between February and May 2026. Two delivered macOS infostealers, and one padded its README with 22 MB of text to beat scanner size limits. OpenClaw banned the accounts and removed the skills.
  • Gateway token theft. NVD describes CVE-2026-25253: before version 2026.1.29, OpenClaw read a gateway URL from a query string, opened a WebSocket to it without asking and sent a token. The CVSS 3.1 score is 8.8, High.
  • Key harvesting. OpenClaw configs and logs hold API keys for model providers and bot tokens. Exposed folders are an easy source of credentials in files.
  • Agent as operator. Because OpenClaw can drive a browser, it can replay a stolen, logged-in session without the user's password or MFA prompt.

For the broader threat picture, see Top Threats to AI Agents.

What Odin Research observed

OpenClaw appears in four Odin Research studies. Hosts are withheld because most belong to developers who made a simple mistake.

  • Leaked agent memory. Monograph 10 found 112 servers exposing hidden agent folders over plain HTTP. OpenClaw folders on 21 servers held 932 secret and session matches, mostly inside live session logs. One session file recorded a cloud metadata query, a Chrome DevTools listing, cloud credentials, a Discord bot token and a reverse shell string. These show what the agent did on its own server, not a breach of anyone else.
  • Resets do not erase secrets. On three OpenClaw servers, resetting a session left backup and reset copies behind. One still held a Google API key. Keys were also copied between agents: one German server kept the same Anthropic key in an OpenClaw profile and a Hermes Agent backup.
  • Key scrapers. Monograph 04 found staged scripts written to find exposed OpenClaw config files through an internet search engine and test the keys against OpenAI, Anthropic and Google. No result on a target was verified.
  • Session cloning. Monograph 24 found an open directory on a server in China that paired Playwright's headless Chromium with an OpenClaw folder. That folder held a complete Chromium profile, with cookie store, Trust Tokens, history and key database, plus the agent's action log. No replay against a real account was seen.
  • Converged workbenches. Monograph 29 found a server in Japan with 1,725 files, including two days of OpenClaw agent logs next to a Hermes backup and a Gemini authentication archive.

How to detect OpenClaw exposure and abuse

  • Check your own surface. Search your external assets for HTTP 200 answers on /.openclaw/ paths. Any hit is a leak to fix today.
  • Watch for folder requests. Requests for agent session logs, auth profiles or model files under /.openclaw/agents/ signal someone harvesting keys.
  • Headless browsers on real profiles. Alert when a headless runtime such as chrome-headless-shell, node or python3 starts with a command line pointing at an OpenClaw browser profile.
  • Skill traffic. Unit 42 recommends comparing a skill's outbound connections with its documentation and flagging undocumented endpoints.

Monograph 10 ships a Suricata signature for requests under the OpenClaw agents folder and a YARA rule for OpenClaw trajectories and auth profiles. Monograph 24 adds Suricata, Sigma, YARA and Sentinel rules for OpenClaw session cloning, and Monograph 04 covers OpenClaw workspace retrieval. All are in the Defense Kit.

How to secure OpenClaw

  1. Patch. Run 2026.1.29 or later to close CVE-2026-25253, and keep Node.js on a release the project recommends.
  2. Keep the Gateway private. Read the project's exposure and sandboxing guides before you expose the Gateway or add other users. Turn on sandboxing for tools.
  3. Deny dotfile paths. Every reverse proxy and file server should return 403 or 404 for /.openclaw/ and other hidden agent folders.
  4. Vet skills. Check publisher provenance and read skill source line by line, as Unit 42 advises. Skills run inside the agent process.
  5. Rotate and purge. When you rotate a key, also delete session backups and reset copies, or the old key stays on disk. See AI Agent Identity and Access.

OpenClaw compared with other AI agents seen in exposed folders

AgentHidden folderWhat it is (Monograph 10)Matches on exposed servers
OpenClaw/.openclaw/Open-source autonomous agent gateway with trajectories and auth profiles932 on 21 servers
Hermes Agent/.hermes/Open-source autonomous agent with multi-profile bots and skill packs967 on 75 servers
Claude Code/.claude/Anthropic command-line coding agent with subagents988 on 26 servers
OpenAI Codex CLI/.codex/OpenAI command-line coding agent with rollout logs74 on 16 servers

Frequently asked questions about OpenClaw

What is OpenClaw?

OpenClaw is an open-source, self-hosted AI assistant that runs on your own computer and takes actions through chat apps such as WhatsApp, Telegram, Slack and Discord, using a local Gateway to manage tools and sessions.

Is OpenClaw malware?

No. OpenClaw is a legitimate MIT-licensed project. The risks come from malicious third-party skills, unpatched versions and exposed folders that leak keys and session logs.

Is OpenClaw safe to use?

It can be used safely with care: keep it updated, keep the Gateway private, enable sandboxing, vet every skill and never serve its .openclaw folder on the web.

Who made OpenClaw?

The repository credits Peter Steinberger and the community. The project is stewarded by the OpenClaw Foundation.

What were Clawdbot and Moltbot?

Earlier names of the same project. NVD's entry for CVE-2026-25253 describes OpenClaw as also known as clawdbot or Moltbot.

What is CVE-2026-25253?

A High-severity OpenClaw flaw, CVSS 3.1 score 8.8. Before version 2026.1.29, OpenClaw took a gateway URL from a query string, opened a WebSocket to it automatically and sent a token.

What are malicious ClawHub skills?

Skills on the ClawHub marketplace that hide harmful code. Unit 42 reported skills that delivered macOS infostealers, injected affiliate links into financial advice and padded files to dodge scanners.

How do I detect OpenClaw leaks on my servers?

Check for HTTP 200 responses on /.openclaw/ paths, watch for requests for agent logs or auth profiles, and flag headless browsers started against OpenClaw profiles. Odin Research publishes rules for these in its Defense Kit.

Sources

  1. OpenClaw, OpenClaw repository, GitHub. github.com/openclaw/openclaw (checked 21 November 2026)
  2. Palo Alto Networks Unit 42, "OpenClaw's Skill Marketplace and the Emerging AI Supply Chain Threat", 23 June 2026. unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/ (checked 21 November 2026)
  3. NVD, CVE-2026-25253. nvd.nist.gov/vuln/detail/CVE-2026-25253 (checked 21 November 2026)

Related tools and terms

All Odin Research explanations · Brandefense glossary

Odin Research

Get the rules, not just the definition

Every Odin Research report ships with verified indicators and detection rules you can load today.

Open the Defense Kit