PricingFAQSign In
🛡️ Security Architecture

Security is not a feature.
It is the foundation.

Cisco found malware in the most popular skills. CrowdStrike published a removal tool. Gartner says: “Not enterprise software.” — The threat is real.

We took everything CrowdStrike, Gartner and Cisco warned about — and fixed it.

🔒 View architecture📋 GDPR & Compliance
Cisco Threat Research: 26% of skills vulnerable
CrowdStrike: Removal tool published
Gartner: “not enterprise software”
The Problem

What happens when AI agents run without security

OpenClaw is powerful — shell access, messaging, web browsing, file system. Without hardening, it’s not just a risk. It’s an open door.

🦠

26% of all skills are vulnerable

Cisco analyzed 31,000 agentic skills. One in four contained security vulnerabilities. The most popular skill was functional malware — with active data exfiltration via curl, prompt injection to bypass safety guidelines, and embedded command injection.

🔴

CrowdStrike publishes removal tool

One of the world’s leading endpoint security vendors classifies OpenClaw as a threat. Their recommendation to enterprises: block and remove. Every IT decision-maker should take note.

⚠️

Gartner: “Not enterprise software”

“There is no quality guarantee, no vendor support, no SLA.” OpenClaw was built for power users — not for enterprises. Without professional management, it’s shadow IT on steroids.

Source: Gartner Analyst Report — “not enterprise software”
💀

Shell + Messaging + Web = Exfiltration Trinity

An OpenClaw agent has simultaneous access to shell commands, web browsing, and messaging. A successful prompt injection attack can exfiltrate data via curl, WhatsApp message, or browser navigation.

OpenClaw Docs: “Running an AI agent with shell access is… spicy.”
Our Model

Shared Responsibility — clearly separated

Like AWS or Azure, we split responsibility transparently. You know exactly what we secure — and what you control.

🟣 ManagedClaw is responsible for

🐳Container IsolationEach customer in their own Docker stack
🔒EncryptionLUKS at rest, TLS in transit
🛡️Egress Filteringnftables, only approved destinations
📡Monitoring & Alerting24/7 surveillance
🔄Updates & PatchesSecurity updates with zero downtime
🔍Skill VettingCisco Scanner + manual review
📝Audit LogsTamper-proof, stored externally
📋GDPR ComplianceDPA, TOMs, deletion policy

🟢 Customer is responsible for

🤖Agent ConfigurationPersonality, tone, rules
📄ContentWhat the agent knows and communicates
🔧Tool SelectionWhich capabilities are enabled
👥Contact ApprovalsWho is allowed to talk to the agent
🔑API KeysManaging your own service credentials
⚖️Usage ComplianceLawful use
Architecture

Technical Security — in detail

No marketing fluff. This is how our security architecture actually works.

🐳

Container Isolation

Each customer runs in their own Docker Compose stack. No shared processes, no shared file system, no shared runtime. One customer cannot see or affect another.

network: "none" — no networking in sandbox capDrop: ["ALL"] — all Linux capabilities removed readOnlyRoot: true — immutable container FS pidsLimit: 256 — fork bomb protection memory: "1g" — OOM isolation user: "1000:1000" — no root
🔒

Encryption

All customer data is double-protected: LUKS full-disk encryption on all Hetzner volumes protects against physical access. TLS 1.3 secures every connection in transit — to the LLM provider, messaging APIs, and to you.

LUKS — full-disk encryption at rest TLS 1.3 — encrypted in transit Per-customer keys — backup encryption isolated Secrets never in prompts — only via env/config
🛡️

Egress Filtering

The biggest vulnerability in self-hosted OpenClaw: an agent can contact any server. With ManagedClaw, the agent can only reach pre-approved destinations. Everything else is blocked at the network level.

nftables — host-level firewall per container DNS filtering — RPZ-based domain blocking Allowlist-only — only WhatsApp, Telegram, LLM API default: DROP — everything else blocked
🔍

Skill Vetting Pipeline

No skill runs without review. Every skill goes through our 5-stage vetting pipeline: source review, automated Cisco Skill Scanner (static + behavioral + LLM analysis), manual code review, sandbox test, and only then: allowlist.

cisco-ai-skill-scanner — AST + behavioral + LLM Commit pinning — exact Git hash, no auto-update Weekly re-scan — weekly repeat review Zero community skills — no deploy without review
🚫

No Elevated Access

OpenClaw has an “elevated mode” that gives the agent host-level access. At ManagedClaw, this mode is permanently disabled — for every customer, every agent, without exception. No agent can break out of its sandbox.

elevated.enabled: false — always, everywhere commands.bash: false — no host shell access commands.config: false — no config access gateway: DENY — agent cannot control gateway
📝

Audit Logs

All agent actions, all admin access, all outbound connections are logged — and streamed to an external, append-only log service. Even we cannot retroactively modify or delete the logs.

Promtail → Loki — centralized log aggregation Append-only — tamper-proof auditd — SSH + sudo fully logged Redaction — API keys automatically masked
🤖

Prompt Injection Defense — Defense in Depth

Prompt injection is an unsolved problem in AI security research. No vendor can claim to have fully solved it — neither can we. But we follow the OpenClaw principle: “Access Control before Intelligence”. We assume the model can be manipulated — and design so that even a successful manipulation causes minimal damage.

Layer 1: Identity
DM pairing + allowlists → only authorized users talk to the agent
Layer 2: Scope
Tool deny lists + sandbox → even with manipulation, reach is limited
Layer 3: Network
Egress filtering → data cannot leave the container
Layer 4: Model
Claude Opus 4.5 (best prompt injection resistance) + content boundaries
Compliance

GDPR & Data Protection

We are a German company, on German servers, under German law. GDPR compliance is not a checkbox for us — it is an obligation.

🇩🇪

Data in Germany

Hosted exclusively on Hetzner servers in Germany (Falkenstein/Nuremberg). No AWS, no Azure, no Google Cloud. No US cloud vendor has access to the infrastructure.

📋

DPA with every customer

Data Processing Agreement per Art. 28 GDPR — with every single customer. Including sub-processor list, TOM reference, and clear deletion policies.

🔧

TOMs documented

Technical and Organizational Measures per Art. 32 GDPR — publicly accessible, not behind a paywall. Because transparency should not be a premium feature.

🗑️

Deletion policy (Art. 17)

Complete data deletion on request — including backups. Documented process with confirmation. We don’t just delete the files — we prove it.

📊

Sub-processor transparency

Open list of all sub-processors: Hetzner (hosting), Anthropic (LLM inference, with DPA + SCCs), log service (EU). Changes are communicated in advance.

🏛️

German company

German UG (limited liability), German law, German jurisdiction. No “we're incorporated in Delaware but say we're GDPR compliant.”

Certification Roadmap

Our path to independently audited security

Now
DPA + TOMs + ROPA
H2 2026
BSI IT-Grundschutz
2027
SOC 2 Type I
2027/28
ISO 27001
Transparency

Trust through evidence

We could read your data.
That’s why we prove that we don’t.

📝

Audit Logs — external & immutable

All access is streamed to an external, append-only log service. Even with root access to our servers, we cannot tamper with the logs. On request, you receive access to the audit logs of your instance.

🔑

SSH fully logged

Every SSH login, every sudo command is captured via auditd and stored externally. We do not routinely access customer systems. If we do: it is documented, traceable, and auditable.

🚫

No routine access

Our management users have no read access to customer data by default. Access requires explicit sudo — which is logged, stored externally, and auditable. No “just taking a quick look.”

📄

TOMs — publicly available

Our Technical and Organizational Measures are not a secret. You can download the document anytime and review it with your DPO.

📥 Download TOMs (PDF)coming soon
📋

DPA template — reviewable upfront

Our Data Processing Agreement is not 80 pages of legalese. Clear, fair, Art. 28 compliant. Fully reviewable before signing.

📥 View DPA template (PDF)coming soon
🔄

Infrastructure as Code

All configurations are stored in a versioned Git repository. Every change has an author, a timestamp, and a reason. No undocumented changes to your infrastructure.

Comparison

DIY vs. Hosted vs. ManagedClaw

Not all OpenClaw offerings are equal. Here’s the difference.

Security FeatureDIY OpenClawHosting Provider*ManagedClaw
Container Isolation No sandbox⚠️ Shared VM Docker per customer, network:none
Egress Filtering Not available Not available nftables allowlist
Skill Vetting No review No review Cisco Scanner + manual
Encryption at rest Plaintext⚠️ Provider-dependent LUKS full-disk
Elevated Access⚠️ Opt-in (often forgotten)⚠️ Unclear Permanently disabled
Audit Logs Local files Not available External, append-only
GDPR Compliance On your own⚠️ Basic DPA + TOMs + deletion policy
Data Location⚠️ Varies⚠️ Often US/Asia Germany (Hetzner)
Prompt Injection Defense No layers No layers 4-layer defense in depth
Support & SLA Community / Discord⚠️ Standard hosting Personal, SLA-backed

* e.g. openclawcloud.work, Tencent Cloud, DigitalOcean — typical “OpenClaw hosting” offerings without security hardening.

Next Step

Ready for an AI agent
you can trust?

Try ManagedClaw — with enterprise security, without the enterprise price. Configuration in 2 minutes, your agent is working within 24 hours.

🔒 DPA-ready · Data in 🇩🇪 · Sandbox isolation from day 1