Skip to main content

Security Overview

OpenClaw is a powerful tool that grants an AI agent significant access to your system. This power comes with real security risks that you must understand and mitigate.

danger

Read this before deploying OpenClaw in any environment with sensitive data. OpenClaw has had critical vulnerabilities, malicious marketplace skills, and tens of thousands of exposed instances discovered in the wild. Gartner characterized it as "a dangerous preview of agentic AI" with "unacceptable cybersecurity risk."


The Security Reality​

OpenClaw's explosive adoption (183k+ GitHub stars, 300k-400k estimated users) massively outpaced its security maturity. In a single week in late January / early February 2026:

  • A one-click RCE vulnerability was discovered and patched (CVE-2026-25253)
  • 40,000-135,000+ exposed instances were found on the public internet
  • 341 malicious skills were discovered on ClawHub
  • A $16 million crypto scam exploited the naming chaos
  • The Moltbook database breach exposed 1.5 million API tokens
  • 93.4% of publicly reachable instances had critical authentication bypasses

As of February 2026, OpenClaw has no bug bounty program and no dedicated security team.


Threat Model​

What OpenClaw Has Access To​

By default, the agent can:

  • Execute arbitrary shell commands as your user
  • Read and write any file your user can access
  • Automate a web browser (fill forms, visit URLs, scrape pages)
  • Send messages on your behalf through connected channels
  • Make HTTP requests to any endpoint
  • Modify its own configuration including SOUL.md (identity/behavior)

Attack Surfaces​

SurfaceRiskMitigation
Gateway WebSocketRemote code execution if exposedBind to localhost, set trustedProxies, enable auth
Reverse proxy bypass93.4% of exposed instances have auth bypassConfigure trustedProxies, overwrite X-Forwarded-For
Channel messagesPrompt injection via incoming messagesContact allowlists, mention-gating
ClawHub skillsMalicious code execution, credential theftVirusTotal scanning, manual review, Clawdex
LLM APIData exfiltration via promptsUse local models for sensitive work
Memory filesSensitive data in plaintextFile permissions, full-disk encryption
SOUL.mdPersistent agent hijacking via prompt injectionClawSec drift detection
CredentialsPlaintext API key storageEnvironment variables, OS keychain, Vault
Browser automationCredential theft, session hijackingDomain allowlists
Tool outputsSensitive data sent to LLM contextSandbox mode, context limits

The Prompt Injection Problem​

OpenClaw processes untrusted content (chat messages, skill outputs, external data) in the same reasoning context as user instructions. There are no hard isolation boundaries between trusted and untrusted content.

Zenity Labs demonstrated a complete zero-click attack chain:

  1. Indirect prompt injection via a Google Document
  2. Agent creates a Telegram backdoor to an attacker-controlled bot
  3. Attacker modifies SOUL.md for persistence across sessions
  4. Creates a scheduled task that re-injects malicious instructions every 2 minutes
  5. Deploys a traditional C2 implant for full system compromise

All of this abuses intended capabilities — no software vulnerability required.


Security Timeline (January-February 2026)​

DateEventSeverity
Jan 27Anthropic sends trademark notice; Steinberger begins rebrand—
Jan 27Crypto scammers hijack X accounts during rebrand, launch $CLAWD pump-and-dump$16M stolen
Jan 30CVE-2026-25253: One-click RCE via gateway URL discovered and patched (v2026.1.29)Critical (8.8)
Jan 31Moltbook database breach — 1.5M API tokens, 35K emails exposedCritical
Feb 2SecurityScorecard: 40,214 exposed instances found via favicon fingerprintingCritical
Feb 2The Register: "security dumpster fire"—
Feb 4Koi Security: 341 malicious ClawHub skills found (ClawHavoc campaign)High
Feb 5Snyk ToxicSkills: 36% of all skills contain security flawsHigh
Feb 5Cisco: 9 vulnerabilities in top community skillHigh
Feb 5Zenity Labs: Prompt injection backdoor research publishedHigh
Feb 7VirusTotal partnership and code safety scanner (v2026.2.6)Mitigation
Feb 8Korean tech firms (Kakao, Naver, Karrot) ban OpenClaw internally—
Feb 9Follow-up scans: 135,000+ exposed instances, 42,665 on ShodanCritical
Feb 9JFrog: 93.4% of exposed instances have auth bypassCritical

Who Has Published Security Research​

OrganizationFinding / Action
Gartner"Unacceptable cybersecurity risk" — recommended blocking OpenClaw downloads
JFrogGateway localhost bypass — 93.4% auth bypass rate
Zenity LabsPrompt injection backdoor via Google Docs, SOUL.md persistence
SecurityScorecard40,214-135,000+ exposed instances via favicon fingerprinting
SnykToxicSkills study — 36% of skills have flaws, 283 leak credentials
Koi Security341 malicious skills, ClawHavoc campaign, AMOS malware
Cisco9 vulnerabilities (2 critical) in top community skill
OX SecurityPlaintext credential storage, credential backup on removal
CrowdStrikeSecurity briefing + Falcon "Search & Removal Content Pack"
BitdefenderTechnical advisory — 900+ malicious skills (~20% of total)
Trend MicroRisk analysis of agentic assistants
Kaspersky"OpenClaw found unsafe for use"
Noma Security53% of enterprise customers gave OpenClaw privileged access in one weekend
Token Security22% of enterprise customers had unauthorized OpenClaw deployments
WizMoltbook database breach discovery
Belgian CCBOfficial advisory urging immediate patching
University of TorontoVulnerability notification for community
Penligent"The security boundary that doesn't exist" — persistence and tool hijack
Cyera"The OpenClaw Security Saga" — how adoption outpaced security

Minimum Security Checklist​

Before using OpenClaw, at minimum:

  • Update to v2026.1.29+ — patches critical RCE
  • Bind gateway to localhost — never expose port 18789
  • Enable authentication — token or password mode
  • Set trustedProxies — if behind any reverse proxy
  • Run openclaw security audit --deep — verify configuration
  • Set channel allowlists — don't accept messages from unknown senders
  • Review installed skills — openclaw skill list and openclaw security scan --all
  • Protect credentials — use environment variables, not plaintext files
  • Set quiet hours — limit autonomous operation
  • Review memory — check ~/.openclaw/memory/ periodically
  • Monitor SOUL.md — watch for unauthorized modifications

Community Security Tools​

ToolStarsDescription
ClawSec265Security skill suite — SOUL.md drift detection, advisory monitoring, audit watchdog
ClawBands34Middleware that enforces human approval before dangerous actions
ClawprintNewTamper-evident audit trail with SHA-256 hash chain ledger
SkillGuardNewSkill file scanner for vulnerabilities and malicious patterns
Security MonitorNew32-script monitoring suite targeting known threat campaigns
openclaw-secureNewHardware-gated secret management with pluggable backends
ClawdexNewPre-installation skill scanning against Koi's malicious skills database

See the Ecosystem page for full details on each tool.


Enterprise Assessment​

"It is not enterprise software. There is no promise of quality, no vendor support, no SLA... it ships without authentication enforced by default." — Gartner

"One of the most dangerous pieces of software a non-expert user can install" — due to explosive viral adoption + deep system privileges + unvetted skills marketplace

For enterprise compliance considerations, see Privacy & Compliance.


Deep Dives​