EVENT DOSSIER
The researchers proposed the CONTINUITY framework, aimed at addressing the issue of security context breaks caused by component boundary crossings in large language model (LLM) agent systems. This framework is based on the assumption of a guaranteed contract model, utilizing mechanisms such as signature root authorization and traceability commitments to carry certified secure contexts across layers. The research team implemented a reference verifier and constructed a deterministic cross-layer fault injection test suite covering 32 types of faults. In 2,560 parameterized attack instances involving 128 fault domain categories, the complete CONTINUITY configuration produced no harmful external effects, successfully completing all 700 benign tasks and upgrading all 200 fuzzy cases.
研究人员提出 CONTINUITY 框架，旨在解决 LLM 代理系统中因组件边界跨越导致的安全上下文中断问题。该框架通过假设 - 保证合同模型各组件，并利用签名根授权、溯源承诺等机制携带经过认证的安全上下文进行跨层传递。研究团队实现了参考验证器及覆盖 32 类故障的确定性跨层故障注入测试套件。在包含 128 个故障域类别的 2,560 个参数化攻击实例中，完整的 CONTINUITY 配置未产生任何有害外部效应，同时完成了所有 700 个良性任务并升级了所有 200 个模糊案例。