clawzero 0.3.0

Deterministic execution boundary for AI agents

ClawZero

Powered by MVAR: https://github.com/mvar-security/mvar

ClawZero is an execution firewall for AI agents.

Blocks CVE-2026-25253, malicious ClawHub skills, and persistent memory injection.
50/50 deterministic block rate in the validation corpus across 9 attack categories. Zero-config API.
Powered by MVAR, the runtime for secure AI agents.

Same input. Same agent. Different execution boundary.
ClawZero enforces policy between model output and tool execution.

Install from PyPI • Documentation • Integration Quickstarts

Quick Start • Why ClawZero • Attack Demo • Witness Artifact

Execution boundary for OpenClaw agents. Powered by MVAR.

pip install clawzero==0.2.1
clawzero doctor openclaw
clawzero demo openclaw --mode compare --scenario shell

Standard OpenClaw -> COMPROMISED
ClawZero -> BLOCKED ✓

Standard OpenClaw executes the attack. ClawZero blocks it deterministically.

ClawZero places a deterministic execution boundary between model output and tool execution.

30-Second Quickstart

pip install clawzero==0.2.1
clawzero doctor openclaw
clawzero demo openclaw --mode compare --scenario shell

Expected output:

Runtime......... OK (mvar-security 1.4.3)
Witness......... OK (chain valid)
Demo............ OK (attack blocked)
Status: SECURE

STANDARD OPENCLAW  →  COMPROMISED
MVAR-PROTECTED     →  BLOCKED ✓
Witness generated  →  YES

OpenClaw Security Crisis

220,000+ exposed instances. 50,000+ RCE-vulnerable via CVE-2026-25253. 1,184 malicious ClawHub skills discovered.

ClawZero blocks exploitation at the execution boundary before credentials leak, before shells execute, and before data exfiltrates.

Addresses 5 of 7 DigitalOcean OpenClaw Security Challenges

Challenge	ClawZero Defense	Command
#1 WebSocket RCE	trusted_websocket_origins checks + exposure diagnostics	clawzero doctor openclaw
#3 Malicious Skills	UNSIGNED_MARKETPLACE_PACKAGE enforcement in prod_locked	clawzero audit decision --profile prod_locked --sink-type tool.custom --target install_skill --package-source clawhub --package-hash sha256:deadbeef --publisher-id unknown-publisher
#4 Credential exfil	Credential sink boundary enforcement	clawzero demo openclaw --mode compare --scenario credentials
#5 Persistent memory	Temporal taint tracking with delayed-trigger enforcement mode	pytest -q tests/test_phaseC_temporal_taint.py
#6 Shadow AI	Witness artifacts + doctor posture checks	clawzero doctor openclaw

Persistent Memory Injection Protection

ClawZero detects delayed-activation attacks where malicious instructions are embedded in agent memory and trigger days later.

Day 1: Agent reads a malicious document.
Day 3: Hidden instruction triggers.
ClawZero: delayed taint reason code path blocks in enforce mode.

Note: We are not aware of other open-source implementations of temporal taint tracking for AI agents.

Reference: tests/test_phaseC_temporal_taint.py.

Enterprise Features

• Compliance-ready audit logs (SARIF export)
• Budget controls (spending limits and abuse detection)
• Package trust validation (blocks unsigned ClawHub skills in prod_locked)
• Network isolation controls (localhost_only / allowlist_only)
• Cryptographically signed witness artifacts
• clawzero doctor openclaw posture check (Status: SECURE)

ClawZero vs Alternatives

Based on public positioning:

• VellaVeto: MCP-specific firewall with formal-verification focus, not OpenClaw-native.
• NemoClaw: NVIDIA managed platform, currently alpha/waitlist.
• Sage: Detection-and-response model that alerts after attempts.
• ClawZero: zero-config runtime enforcement, IFC taint-aware policy, production-ready today.

Decision shortcuts:

• Need formal-verification-first MCP posture: VellaVeto.
• Need managed platform lifecycle: NemoClaw (when GA).
• Need detection + response workflow: Sage.
• Need zero-config execution firewall: ClawZero.

Adapters

OpenClaw adapter is included and works out of the box:

pip install clawzero

LangChain adapter code is included, and requires LangChain packages in your project:

pip install clawzero langchain langchain-openai

LangChain Integration

from clawzero.adapters.langchain import protect_langchain_tool

safe_tool = protect_langchain_tool(
    my_langchain_tool,
    sink="filesystem.read",
    profile="prod_locked",
)

Run the packaged example:

python examples/langchain_integration.py

Protect Entire Agents

from clawzero import protect_agent

safe_agent = protect_agent(agent, profile="prod_locked")

protect_agent() auto-detects common framework patterns and wraps registered tools with deterministic sink enforcement.

Additional Framework Adapters

CrewAI and AutoGen adapters are now included alongside OpenClaw and LangChain:

from clawzero.adapters.crewai import protect_crewai_tool
from clawzero.adapters.autogen import protect_autogen_function

MCP adapter alpha is now available:

from clawzero.adapters.mcp import MCPAdapter

adapter = MCPAdapter(profile="prod_locked")
safe_client = adapter.wrap_client(mcp_client, method_name="call_tool")

Attack Pack Validation (50 Vectors)

Run the packaged attack corpus:

pytest tests/attack_pack/ -v

Categories covered: command injection, path traversal, credential exfiltration, data exfiltration, persistence, lateral movement, supply chain, social engineering, and denial of service.

Benchmark

Measure policy decision latency:

python -m clawzero.benchmark --iterations 1000

This reports per-scenario mean/p95/p99 latency and throughput for deterministic sink enforcement.

Why ClawZero?

Autonomous AI agents frequently execute tool calls with high privileges.

When these agents ingest untrusted input, prompt injection can escalate into:

• shell execution
• filesystem access
• credential leakage
• data exfiltration

ClawZero prevents these escalations by enforcing deterministic policy checks at execution sinks before commands run.

Threat Model

OpenClaw agents commonly run with tools capable of:

• shell execution
• filesystem access
• credential retrieval
• outbound network requests

When these agents process untrusted documents or user input, hidden instructions can influence tool calls.

Without an execution boundary, these instructions can trigger high-privilege operations.

ClawZero intercepts these tool calls and enforces policy before execution occurs.

Attack Demo Proof

The attack demo exists to demonstrate runtime enforcement behavior.

ClawZero is not a model safety claim.

It is an execution boundary claim.

The demo illustrates how untrusted input can influence agent tool calls and how the ClawZero boundary blocks those actions deterministically.

Run the side-by-side comparison:

clawzero demo openclaw --mode compare --scenario shell
clawzero demo openclaw --mode compare --scenario credentials
clawzero demo openclaw --mode compare --scenario benign

Session + Wrap (v1)

wrap v1 enforces at the process/tool call boundary. It does not claim syscall-level interception.

clawzero wrap -- python my_agent.py
clawzero session status <session_id>
clawzero session report <session_id> --format json
clawzero keys show

Security and Responsible Use

ClawZero is a defensive security component designed to enforce execution boundaries for AI agents.

The project includes attack demonstrations and adversarial scenarios to show how prompt injection and untrusted inputs can reach high-privilege execution sinks.

These demonstrations exist solely for defensive research and education.

When using ClawZero or its demonstrations:

• Only test systems you own or have explicit authorization to evaluate
• Run demonstrations in sandboxed or isolated environments
• Treat automated results as signals; verify findings manually

ClawZero is designed to prevent exploitation, not enable it.

The attack demonstrations show how enforcement works; they are not tools for performing real-world attacks.

Canonical Witness Artifact

{
  "timestamp": "2026-03-12T10:00:00Z",
  "agent_runtime": "openclaw",
  "sink_type": "shell.exec",
  "target": "bash",
  "decision": "block",
  "reason_code": "UNTRUSTED_TO_CRITICAL_SINK",
  "policy_id": "mvar-security.v1.4.3",
  "engine": "mvar-security",
  "provenance": {
    "source": "external_document",
    "taint_level": "untrusted",
    "source_chain": ["external_document", "openclaw_tool_call"],
    "taint_markers": ["prompt_injection", "external_content"]
  },
  "adapter": {
    "name": "openclaw",
    "mode": "event_intercept",
    "framework": "openclaw"
  },
  "witness_signature": "ed25519:d91fd8f73f3d05f8ec7b3d8e5e7cf2e27869a5f0f1ee3bd17da2df5ec41c9cb2a3c7e4f3540b4f7f4f948f0f185318273447bcb0adf24a4b2a1b53b7a1b2c90a"
}

What ClawZero Is / Is Not

ClawZero is:

• an in-path runtime enforcement substrate
• deterministic sink policy evaluation
• a signed witness artifact generator

ClawZero is not:

• a red-team toolkit
• an attack simulation platform
• an LLM-as-judge safety layer

CLI

Command families map to enforcement jobs:

• clawzero prove - run install-to-proof validation in one command
• clawzero demo - run side-by-side enforcement proof demos
• clawzero witness - inspect and validate witness artifacts
• clawzero audit - evaluate deterministic decisions for sink requests
• clawzero attack - replay known attack scenarios as enforcement proofs
• clawzero report - export witness artifacts to SARIF for code scanning

Note: clawzero prove is available on the main branch and will be included in the next PyPI release.

Zero-Config API

from clawzero import protect

safe_tool = protect(
    my_tool,
    sink="filesystem.read",
    profile="prod_locked"
)

Policy Profiles

Sink Type	dev_balanced	dev_strict	prod_locked
shell.exec	block	block	block
filesystem.read	allow, block /etc/**, ~/.ssh/**	block, allow /workspace/**	block, allow /workspace/project/**
filesystem.write	allow, block /etc/**, ~/.ssh/**	block, allow /workspace/**	block, allow /workspace/project/**
credentials.access	block	block	block
http.request	allow	allow mode + block all domains	allow mode + allow localhost
tool.custom	allow	annotate	allow

Powered by MVAR

MVAR is the enforcement engine. ClawZero is the OpenClaw adapter. MVAR governs the sink policy enforcement decisions.

• MVAR repository: https://github.com/mvar-security/mvar
• Filed as provisional patent (February 24, 2026, 24 claims)
• Submitted to NIST RFI Docket NIST-2025-0035
• Published as preprint on SSRN (February 2026)

Early Release - Join Us

This is early. The clawzero demo shows enforcement in harness + OpenClaw simulation.

Real multi-turn agent testing is next.

If you're running agents (LangChain, CrewAI, AutoGen, OpenClaw, etc.) and want to try it live:

• DM @Shawndcohen on X
• Open an issue with your setup/framework

Happy to pair debug and share results.

License

Apache 2.0