arXiv: Systematic Security Evaluation of 6 OpenClaw Variants — All Show 'Substantial Security Risk'

arXiv·medium signal

Paper (2604.03131) evaluates six Claw-series agents (OpenClaw, AutoClaw, QClaw, KimiClaw, MaxClaw, ArkClaw) across 205 test cases spanning the full agent execution lifecycle. Every agent exhibited substantial security risk, with overall system risk significantly higher than the underlying model alone. Average defense rate was only 17% across 47 adversarial scenarios. 26% of community-contributed tools contain vulnerabilities. Builders deploying any OpenClaw variant should treat this as a mandatory security review.