open/claude-code-complete-agent
1
0
Fork 0
Code Activity Actions

docs(security): add OpenClaw vs Claude Code security assessment

Browse source 
Data-driven comparison covering 9 CVEs, 10 security categories,
and attack surface analysis. Based on published research from
SecurityScorecard, DigitalOcean, Sangfor, and OpenClaw official docs.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Kjell Tore Guttormsen 2026-04-05 23:38:34 +02:00
commit 2fe6a78e3c
3 changed files with 316 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

1
security/README.md
View file

@ -14,3 +14,4 @@ agent behavior.
| `auto-mode-explained.md` | Auto Mode: AI safety classifier for autonomous execution |
| `hook-based-guardrails.md` | Building custom security with PreToolUse hooks |
| `nemoclaw-comparison.md` | Honest comparison with NemoClaw's enterprise security |
| `openclaw-security-assessment.md` | Data-driven security comparison: OpenClaw CVEs vs Claude Code architecture |

300
security/openclaw-security-assessment.md Normal file
View file

@ -0,0 +1,300 @@
# OpenClaw vs Claude Code: Security Assessment
A data-driven comparison of security posture between OpenClaw
(self-hosted AI agent) and Claude Code (managed AI agent). All
CVE numbers, exposure statistics, and incident references are
sourced from published security research (March-April 2026).
This is not a marketing document. Both platforms have strengths
and gaps. The goal is to help you choose based on evidence.
## The security landscape (March 2026)
OpenClaw reached 247K GitHub stars and became the fastest-growing
open-source project in history. With adoption came scrutiny:
- **9 CVEs in 4 days** (March 18-21, 2026), CVSS scores up to 9.9
- **40,214 internet-exposed instances** (SecurityScorecard)
- **35-63% of deployments vulnerable** at time of analysis
- **824 malicious skills** found in ClawHub marketplace out of 10,700+
- **128 pending CVE assignments** in upstream tracker
Claude Code has no public CVEs to date. Its managed infrastructure
model eliminates several attack surface categories entirely.
## OpenClaw vulnerability summary
| CVE | CVSS | Category | Impact |
|-----|------|----------|--------|
| CVE-2026-22172 | 9.9 | Auth bypass | Client self-declares admin scope |
| CVE-2026-25253 | 8.8 | WebSocket hijack | One-click RCE via malicious link |
| CVE-2026-32048 | 7.5 | Sandbox escape | Child processes spawn unsandboxed |
| CVE-2026-32025 | 7.5 | Brute force | No rate limiting on localhost auth |
| CVE-2026-32032 | 7.0 | Shell injection | Untrusted SHELL env variable |
| CVE-2026-29607 | 6.4 | Approval bypass | Approve safe command, swap payload |
| CVE-2026-28460 | 5.9 | Allowlist evasion | Line-continuation character bypass |
| CVE-2026-22171 | 8.2 | Path traversal | Arbitrary file write via media |
| CVE-2026-32049 | 7.5 | DoS | Oversized media payload |
**Architectural patterns in these CVEs:**
1. Authorization model accepts client-declared permissions (CVE-2026-22172)
2. Sandbox constraints don't propagate through process spawning (CVE-2026-32048)
3. Human-in-the-loop approvals have enforcement gaps (CVE-2026-29607, 28460)
4. Rate limiting absent on authentication endpoints (CVE-2026-32025)
## Head-to-head: 10 security categories
### 1. Network exposure
**OpenClaw risk:** Gateway on port 18789. Default binds to localhost
but misconfigured reverse proxies, Docker port publishing, and
Tailscale Serve can expose it. 40K+ instances found on public
internet. The gateway is a WebSocket server — no origin validation
led to CVE-2026-25253 (one-click RCE from any website).
**Claude Code:** No gateway. No listening port. No WebSocket server.
The attack surface does not exist. Claude Code communicates with
Anthropic's API over outbound HTTPS only.
**Verdict:** Claude Code eliminates this category entirely.
### 2. Authentication and access control
**OpenClaw risk:** Token/password/trusted-proxy modes. Gateway
token required by default since 2026.1.29, but older deployments
lack auth. CVE-2026-22172 allowed any authenticated user to become
admin by self-declaring scope. DM pairing (1-hour expiring codes)
provides messaging access control.
**Claude Code:** Single-user model. No multi-user auth layer needed.
Permission modes (default, auto-edit, auto, bypass) control what
the agent can do, not who can access it. API key stored in OS
keychain (macOS) or environment variable.
**Verdict:** Different models. OpenClaw needs auth because it's
multi-user and network-accessible. Claude Code is single-user and
local, so the auth question doesn't arise. For multi-user needs,
OpenClaw must be properly configured; Claude Code isn't designed
for it.
### 3. Execution sandboxing
**OpenClaw risk:** Three-level sandbox (off, non-main, all). Docker
containerization available. But CVE-2026-32048 showed sandbox
constraints don't propagate to child processes. NemoClaw adds
kernel-level enforcement (Landlock, seccomp, netns).
**Claude Code:** Permission modes + hooks. macOS sandbox-exec
available. Hooks run as separate processes and can block any tool
call. No kernel-level isolation by default, but the agent is
prevented from attempting dangerous operations rather than
contained after attempting them.
**Verdict:** NemoClaw wins for enterprise isolation. Claude Code
wins for flexibility and zero infrastructure. Vanilla OpenClaw has
known sandbox escape paths.
### 4. Supply chain (skills/plugins)
**OpenClaw risk:** ClawHub marketplace had 824 malicious skills
among 10,700+ (the ClawHavoc campaign). Atomic macOS Stealer
distributed via fake skills. Publishing requires only a week-old
GitHub account with no code review.
**Claude Code:** Plugin marketplace is smaller (2,300+) with a
review process. Local plugins don't go through any marketplace.
The llm-security plugin provides supply chain scanning: blocklists
for 7 package managers, OSV.dev CVE checking, Levenshtein
typosquat detection, and npm/pip audit integration.
**Verdict:** Claude Code's ecosystem is smaller but more controlled.
OpenClaw's marketplace scale introduced real malware distribution.
### 5. Prompt injection defense
**OpenClaw risk:** Prompt injection led to code execution
(CVE-2026-30741). Persistent memory creates delayed-execution
attack paths: malicious instructions embedded in documents can
remain dormant for days. Official guidance: "validate tool calls
against policy, not only model output."
**Claude Code:** llm-security provides 3 layers of defense:
- `pre-prompt-inject-scan.mjs`: Blocks injection patterns in
user prompts (configurable: block/warn/off)
- `post-mcp-verify.mjs`: Scans ALL tool output for injection
and HTML content traps
- `post-session-guard.mjs`: Detects runtime trifecta patterns
(untrusted input + sensitive data + exfiltration sink)
Additionally, the taint-tracer scanner traces data flow paths
statically, and the toxic-flow-analyzer correlates findings
across scanners to detect compound attack chains.
**Verdict:** Claude Code (with llm-security) has more active
defense layers. OpenClaw relies on tool policy enforcement which
has documented bypass paths (CVE-2026-29607, 28460).
### 6. Credential management
**OpenClaw risk:** API keys and platform tokens stored in plaintext
in `~/.openclaw/`. File permissions (600/700) are the primary
protection. Prompt injection attacks can exfiltrate credentials
through tool calls.
**Claude Code:** API key stored in macOS Keychain (encrypted,
OS-level). llm-security hooks block credential patterns in file
writes (`pre-edit-secrets.mjs`) and detect secrets in code
(`entropy-scanner`). Path guard blocks writes to `.env`, `.ssh/`,
`.aws/`, and credentials files.
**Verdict:** Claude Code's approach (OS keychain + write blocking)
is stronger than filesystem permissions on plaintext files.
### 7. Browser and relay security
**OpenClaw risk:** Browser profiles with logged-in sessions become
agent-accessible. Remote CDP connections, SSRF via browser (private
network access enabled by default). Relay access should be
restricted to approved operators.
**Claude Code:** Playwright MCP for browser automation. Computer
Use in Desktop app. No persistent browser relay. No agent access
to browser profiles by default. Each browser session is explicit.
**Verdict:** Claude Code has a smaller browser attack surface
because there is no always-on relay.
### 8. Session isolation
**OpenClaw risk:** Multi-user access requires explicit session
isolation (`dmScope: per-channel-peer`). Default is unified session,
creating cross-user context leakage risk. Any allowed sender can
induce tool calls within the agent's permission set.
**Claude Code:** Single-user, single-session model. No cross-user
leakage possible. Agent Teams run in isolated contexts (separate
worktrees for file isolation).
**Verdict:** Claude Code eliminates multi-user leakage by design.
### 9. Configuration hardening
**OpenClaw risk:** Numerous configuration surfaces (openclaw.json,
SOUL.md, tool policies, network binding, mDNS discovery, DM
policies). Default configurations have historically been too
permissive. `openclaw security audit --deep` provides automated
checking.
**Claude Code:** Configuration through settings.json hierarchy
(global, project, local) and CLAUDE.md. The config-audit plugin
analyzes configuration quality with A-F grading. The
reference-config-generator creates hardened configurations based
on detected gaps.
**Verdict:** Both have audit tooling. OpenClaw has more
configuration surface area (and more ways to misconfigure).
### 10. Monitoring and incident response
**OpenClaw risk:** Logging to `/tmp/openclaw/`. Redaction available
but opt-in. No built-in anomaly detection. Audit trail integrity
not guaranteed (file-based logs).
**Claude Code:** llm-security provides runtime monitoring:
`post-session-guard.mjs` tracks tool call patterns in a sliding
window (20 calls), detects concentrated MCP usage, and tracks
cumulative data volume (100KB/500KB/1MB thresholds). The dashboard
aggregator provides cross-project posture visibility. Scan
baselines enable drift detection over time.
**Verdict:** Claude Code (with llm-security) has more active
runtime monitoring. Both lack tamper-resistant audit trails.
## What Claude Code does NOT do
Honest gaps where OpenClaw has genuine advantages:
| Capability | OpenClaw | Claude Code |
|------------|----------|-------------|
| Daemon persistence | Runs 24/7 as background process | Session-based, stops when closed |
| Multi-engine | Claude, GPT, Gemini, local models | Claude only |
| Native messaging | 15+ channels (WhatsApp, Telegram, Signal, iMessage) | Channels (limited), MCP bridges |
| Canvas/A2UI | Interactive HTML workspace | HTML generation only |
| Self-hosting | Full infrastructure control | Anthropic-dependent |
| Kernel isolation | Via NemoClaw (Landlock, seccomp) | Not available |
These gaps matter for specific use cases. If you need always-on
daemon persistence or kernel-level multi-tenant isolation, Claude
Code is not a drop-in replacement.
## The "use Claude Code" mitigation
For use cases where Claude Code covers the functional requirements
(21 of 22 capabilities — see feature-map.md), migrating from
OpenClaw to Claude Code eliminates entire attack surface categories:
| OpenClaw vulnerability | Claude Code mitigation |
|----------------------|----------------------|
| WebSocket hijacking (CVE-2026-25253) | No gateway, no listening port |
| 40K exposed instances | No network exposure |
| ClawHub malware (824 malicious skills) | Local plugins, smaller reviewed marketplace |
| Plaintext credential storage | OS keychain encryption |
| Prompt injection to RCE (CVE-2026-30741) | Multi-layer hook defense |
| Sandbox escape (CVE-2026-32048) | Permission-based prevention |
| Approval bypass (CVE-2026-29607) | Deterministic hook validation |
| Auth bypass (CVE-2026-22172) | Single-user, no multi-user auth |
| Shadow AI (22% enterprise) | Anthropic billing visibility |
| Moltbook breach (2.8M agents) | No shared agent platform |
This is not "Claude Code is more secure" as a blanket claim. It is
"Claude Code's architecture does not have these specific attack
surfaces." The tradeoff is less infrastructure control and vendor
dependency.
## Security tooling comparison
| Tool | OpenClaw | Claude Code (with llm-security) |
|------|----------|-------------------------------|
| Built-in audit | `openclaw security audit --deep` | `/security posture` + `/security audit` |
| Prompt injection defense | Tool policy validation | 3 active hooks + static taint analysis |
| Supply chain scanning | Manual review | 7 package managers, OSV.dev, typosquat detection |
| Secret detection | None built-in | Entropy scanner + write-blocking hooks |
| Memory poisoning | None built-in | Memory poisoning scanner (CLAUDE.md, rules) |
| Attack simulation | None built-in | 38 scenarios across 7 categories |
| Continuous monitoring | None built-in | `/security watch` with baseline diffing |
| Threat modeling | None built-in | Interactive STRIDE/MAESTRO sessions |
| Framework coverage | Internal checks | OWASP LLM Top 10, Agentic AI Top 10, Skills Top 10, MCP Top 10, AI Agent Traps |
## Recommendation
**Choose based on your actual requirements:**
1. **Personal automation, development work:** Claude Code. Smaller
attack surface, active hook defense, no infrastructure to secure.
2. **Always-on daemon, multi-channel messaging:** OpenClaw with
hardened configuration. Follow the Blink 10-step checklist.
Consider NemoClaw for kernel isolation.
3. **Enterprise, multi-tenant, compliance:** NemoClaw on OpenClaw
or wait for Claude Code enterprise features. Neither vanilla
OpenClaw nor Claude Code currently meets SOC2 requirements.
4. **Maximum security on personal setup:** Claude Code + llm-security
plugin. This repo demonstrates the configuration.
## Sources
All vulnerability data sourced from published research:
- OpenClaw CVE flood analysis (openclawai.io, March 2026)
- SecurityScorecard exposure report (40,214 instances)
- DigitalOcean "7 OpenClaw Security Challenges" (2026)
- Sangfor "OpenClaw Security Risks" analysis
- OpenClaw official security documentation (docs.openclaw.ai)
- Valletta Software hardening guide (2026)
- Nebius architecture analysis
CVE numbers verified against NVD. Statistics represent point-in-time
measurements from the cited reports.