Skip to Content
OpenclawOpenClaw Security — Sandbox Isolation, CVE Patching & Threat Protection

OpenClaw Security

OpenClaw’s security track record in early 2026 has been challenging. Microsoft, CrowdStrike, Kaspersky, and Sophos have all published advisories. Here’s what happened and how Curate-Me addresses every known attack vector.

The Security Landscape

EventDateSeveritySource
CVE-2026-25253 — One-click RCE via WebSocket hijackingJan 2026CVSS 8.8NVD
512 vulnerabilities in security audit (8 critical)Jan 2026CriticalIndependent audit
341 malicious ClawHub skills (ClawHavoc campaign)Feb 2026HighSecurity researchers
42,665 publicly exposed instancesFeb 2026HighShodan scan
OpenClaw creator joins OpenAI — project to foundationFeb 14, 2026Governance riskOfficial announcement

Our Security Architecture

Layer 1: Sandbox Isolation

Every agent session runs in an isolated sandbox with one of four access tiers:

TierFilesystemNetworkUse Case
READ_ONLYRead project files, write only /tmpNone during executionAnalysis, reporting
WRITE_RESTRICTEDExplicit path patterns onlyAllowlisted domains onlyControlled automation
WRITE_PROJECTAnywhere in project treeBridge networkDevelopment tasks
FULL_ACCESSUnrestrictedFull networkTrusted admin only

Deny patterns (always blocked, regardless of tier):

  • .env, .env.* — Environment files with secrets
  • *.pem, *.key — Cryptographic keys
  • .git/config — Git credentials
  • ~/.ssh/* — SSH keys
  • ~/.aws/* — AWS credentials

Per-session byte-level write tracking with configurable size limits. Immutable paths are never writable regardless of tier.

Layer 2: Network Phase Separation

Agent execution is divided into phases with different network access:

PhaseNetworkPurpose
SETUPONInstall dependencies, git clone, download models
EXECUTIONOFFAgent runs locally — no outbound connections
TEARDOWNONUpload artifacts, push results, send notifications

During the EXECUTION phase, agents cannot make outbound network calls. This prevents:

  • Data exfiltration to attacker-controlled servers
  • C2 (command & control) communication
  • Cryptocurrency mining pool connections
  • Unauthorized API calls

Four access levels per phase: FULL, ALLOWLIST_ONLY, LOCAL_ONLY, NONE. Domain and port filtering available in allowlist mode.

Layer 3: PII & Secrets Scanning

Every LLM request is scanned for sensitive content before it reaches the provider:

14 built-in patterns:

  • Social Security Numbers (SSN)
  • Credit card numbers (Luhn validation)
  • API keys (OpenAI, Anthropic, AWS, Stripe, etc.)
  • Email addresses
  • Phone numbers
  • IBAN / VAT numbers (EU compliance)
  • Passport numbers
  • Medication dosages (EU AI Act)

Presidio NER integration (feature-flagged):

  • Named entity recognition for PII that regex misses
  • Person names, addresses, medical records
  • Configurable confidence threshold

When PII is detected:

  • allow — Log finding, forward request (audit trail only)
  • block — Deny request, return governance error
  • needs_approval — Queue for human review (HITL)

Layer 4: Human-in-the-Loop (HITL)

High-cost or sensitive operations require human approval:

TriggerDefault ThresholdConfigurable
Cost > threshold$10/requestYes
External API callsAlwaysYes
File deletionsAlwaysYes
Git pushAlwaysYes
New skill installationCAUTION-ratedYes

Approval queues visible in the dashboard with SLA tracking. Auto-deny after configurable timeout (default: 24 hours).

CVE Protection

Curate-Me tracks all known OpenClaw CVEs and enforces a minimum version floor:

MIN_SUPPORTED_VERSION = "2026.3.7" # Updated after each CVE disclosure

Runners below the minimum version:

  1. Cannot start new sessions
  2. Show CVE exposure warnings in the dashboard
  3. Are flagged for auto-upgrade in the next maintenance window

The Security Posture dashboard (/runners/security) shows:

  • CVE exposure per runner
  • CVSS scores and severity
  • Affected vs patched status
  • One-click auto-patch button
  • Remediation history

Incident Prevention Map

Known AttackHow We Block It
CVE-2026-25253 (RCE)Network phase separation blocks outbound during execution
Docker sandbox escape4-tier sandbox with deny patterns prevents host access
ClawHavoc malicious skillsClawHavoc-hardened scanner with VirusTotal + YARA rules
42,665 exposed instancesAgents run behind gateway auth — zero public surface
Runaway agent purchasesHITL approval queues for high-cost operations
API key / PII leakagePII scanning intercepts secrets before they reach providers

Security Score

The Security Posture dashboard calculates an aggregate score (0-100) based on:

  • CVE exposure (are all runners patched?)
  • Skill scan coverage (are all installed skills scanned?)
  • Policy compliance (are governance policies configured?)
  • Patch currency (is the OpenClaw version current?)

Trend tracking over 30 days shows whether your security posture is improving.

Compliance

  • Immutable audit trail for every agent action
  • Export formats: JSON, PDF
  • SOC 2 compatible log structure (Enterprise plan)
  • GDPR: Data residency options, PII scanning, right to deletion
  • EU AI Act: Medication dosage scanning, VAT/IBAN detection

Further Reading