ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Workflow Analysis
arXiv:2603.24414
Updated Oct 2026 · CipherClaw v1.1.0
Open-source observability for multi-agent AI and published research on OpenClaw security, from ClawKeeper (arXiv:2603.24414) to the 2026 Claw Chain disclosures.
2026 research log
Agents take new prompts, call tools, and cross runtime boundaries every minute. This year's work shows the loop clearly: publish, get attacked, patch, measure, repeat.
Feb 2026
OpenClaw SKILL.md, CI, examples, and core capability tests published.
Mar 2026
Three-layer real-time protection (skills, plugins, workflow) and a zero-fork gateway security layer.
Mar 2026
OpenTelemetry converter, publish-ready package, ESM + CJS builds.
Apr 2026
Six OpenClaw-family frameworks tested; autonomous retries create a cost paradox under complex attacks.
Apr 2026
Four chained CVEs enabling sandbox escape disclosed Apr 22 and patched the next day.
Jun 2026
Practical mitigation guide and Skill for OpenClaw users (arXiv:2606.11007).
Oct 2026
Vulnerable test tooling updated; Node 18 runtime checks preserved; 235 tests green.
Benchmarks & evidence
Results are stated within the scope each source tested. None of them claims to be a complete defense.
| Measure | Artifact | Interpretation |
|---|---|---|
| Threat mitigation | ClawKeeper (arXiv:2603.24414) | Outperformed listed open-source baselines in the tested setup |
| Variant evaluation | 6 OpenClaw-family frameworks (arXiv:2604.03131) | Autonomous recovery can make failed attacks costlier than successful ones |
| Disclosure response | Claw Chain, 4 CVEs (CSA) | Disclosed Apr 22, 2026; patched within one day |
| Tooling quality | CipherClaw v1.1.0 | 235/235 tests passing, zero type errors |
01 · Real-time protection research
A three-layer architecture (skills, plugins, workflow analysis) using decoupled system-level middleware to continuously monitor and intervene in OpenClaw agent behavior. In benchmark testing it outperformed existing open-source baselines.
git clone https://github.com/Alexi5000/ClawKeeper.git
cd ClawKeeper
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txtpython examples/run_clawkeeper_eval.py --config configs/openclaw-baseline.yaml02 · OpenClaw debug agent · v1.1.0
Open-source debugging and observability for multi-agent systems: trace root causes across agent chains, profile behavior, flag anomalies, and export traces through the OpenTelemetry converter. Ships as an OpenClaw skill with 235 passing tests.
git clone https://github.com/Alexi5000/CipherClaw.git
cd CipherClaw
npm install
npm test # 235 testsnpx tsx examples/openclaw-integration.ts
# attach an OpenClaw trace source, export via the OTEL converter03 · Deployment surface
The self-hosted agent runtime this work instruments: container boundaries, gateway behavior, and skill integration. 2026 disclosures like Claw Chain show why runtime security has to be continuous.
git clone https://github.com/openclaw/openclaw.git
cd openclaw
# follow the repository README for runtime-specific setup# keep OpenClaw patched past the Apr 2026 Claw Chain fixes
openclaw --config ./configs/local.yamlResearch papers
Every claim on this page links back to one of these. Open the source or download the PDF.
arXiv:2603.24414
arXiv:2603.11853
arXiv:2604.03131
arXiv:2606.11007
Cloud Security Alliance research note
Production Triage gives your team a scoped review: architecture diagrams, failure modes, and patch posture, all tied to sources.
Book Production Triage at TechTide AI