Preprint reports multi-stage security gateway for coding agents

An arXiv preprint describes ClawSentry, an open-source security supervision gateway for agent runtimes that reviews skills before use and applies progressive checks during and after tool activity. The authors report that its deterministic, semantic-review and read-only evidence-seeking layers reduced contextual attack-success rates in their SkillInject tests while retaining clean-skill task success. The system also uses session-level checks intended to recognize tool switching and rephrased retries. Those results are preprint claims, not independent validation of production protection.