跳转至

CONTINUITY:面向可组合大模型智能体控制的安全上下文契约

文章背景与核心概要

随着大语言模型(LLM)智能体系统日益复杂,它们越来越多地集成了各种不同的安全机制——例如溯源追踪、授权引擎、策略强制监视器、协议适配器以及执行控制。然而,行业内存在一个关键的漏洞:各自独立的正确安全机制,不一定能组合成端到端安全的系统。当安全攸关的上下文跨越组件边界时,它们可能会被丢弃、放宽、重新绑定或误解。本文识别并形式化了这种失效模式,将其命名为安全上下文不连续性(security-context discontinuity),并推出了 CONTINUITY——一个旨在实现智能体安全控制可验证组合的全新框架。

CONTINUITY 通过使用假设-保证契约(assume-guarantee contract)对每个独立的系统组件进行建模,从而解决了安全上下文不连续的问题。它利用一套严密的密码学原语,在组件转换过程中维护经过身份验证的安全上下文,包括签名根授权、溯源承诺、绑定角色的转换回执等。评估结果表明,在包含 2,560 个参数化攻击实例的测试中,完整配置的 CONTINUITY 在攻击条件下实现了零有害外部副作用,同时在良性任务上保持了 100% 的成功率。


1. 执行摘要 (Executive Summary)

随着大语言模型(LLM)智能体系统日益复杂,它们越来越多地集成了各种不同的安全机制——例如溯源追踪、授权引擎、策略强制监视器、协议适配器以及执行控制。然而,行业内存在一个关键的漏洞:各自独立的正确安全机制,不一定能组合成端到端安全的系统

As Large Language Model (LLM) agent systems grow more complex, they increasingly integrate disparate security mechanisms—such as provenance tracking, authorization engines, policy enforcement monitors, protocol adapters, and execution controls. However, a critical gap exists: individually correct security mechanisms do not necessarily compose into an end-to-end secure system.

当安全攸关的上下文跨越组件边界时,它们可能会被丢弃、放宽、重新绑定或误解。本文识别并形式化了这种失效模式,将其命名为安全上下文不连续性(security-context discontinuity),并推出了 CONTINUITY——一个旨在实现智能体安全控制可验证组合的全新框架。

When security-critical contexts cross component boundaries, they can be dropped, widened, rebound, or misinterpreted. This paper identifies and formalizes this failure mode as security-context discontinuity and introduces CONTINUITY, a novel framework designed to achieve verifiable composition of agent security controls.


2. 核心概念与架构 (Key Concepts & Architecture)

CONTINUITY 通过使用假设-保证契约(assume-guarantee contract)对每个独立的系统组件进行建模,从而解决了安全上下文不连续的问题。它利用一套严密的密码学原语,在组件转换过程中维护经过身份验证的安全上下文:

CONTINUITY addresses security-context discontinuity by modeling each individual system component with an assume-guarantee contract. It maintains authenticated security context across transitions utilizing a rigorous suite of primitives:

  • 签名根授权(Signed Root Grants): 建立初始信任根。
  • 溯源承诺(Provenance Commitments): 追踪数据和指令的起源与演变。
  • 绑定角色的转换回执(Role-Bound Transition Receipts): 确保权限在交接过程中与智能体角色紧密耦合。
  • 有界类型释放(Bounded Typed Releases): 在严格的类型约束下安全地降级或暴露数据。
  • 转换见证(Transformation Witnesses): 密码学验证中间转换是否保留了安全不变量。
  • 效果绑定的执行许可(Effect-Bound Execution Permits): 授予执行外部操作的精确且可撤销的权限。
  • Signed Root Grants: Establishing initial trust roots.
  • Provenance Commitments: Tracking the origin and evolution of data/instructions.
  • Role-Bound Transition Receipts: Ensuring permissions are tightly coupled to agent roles during handoffs.
  • Bounded Typed Releases: Safely downgrading or exposing data under strict type constraints.
  • Transformation Witnesses: Cryptographically verifying that intermediate transformations preserve security invariants.
  • Effect-Bound Execution Permits: Granting precise, revocable authority to execute external actions.

通过这些机制,该框架将端到端结果完整性(end-to-end consequence integrity)形式化,要求每一个已实现的外部效果都必须有有效且当前授权见证的支持,该见证需将主体、任务、溯源、委托、策略状态、规范动作以及终结边界紧密联系在一起。

Through these mechanisms, the framework formalizes end-to-end consequence integrity, requiring that every realized external effect is backed by a valid, current authorization witness linking the principal, task, provenance, delegation, policy state, canonical action, and finality boundary.


3. 评估与结果 (Evaluation & Results)

为了验证该框架,作者实现了一个参考验证器,以及一个确定性的跨层故障注入套件,涵盖了四个独特应用域中的 32 个不同故障类别。

To validate the framework, the authors implemented a reference verifier alongside a deterministic cross-layer fault-injection suite covering 32 distinct fault classes across four unique application domains.

  • 测试规模(Test Scale): 评估跨越了 2,560 个参数化攻击实例,涵盖 128 个故障域类别。
  • 良性性能(Benign Performance): 成功完成了 所有 700 个良性任务(100%),且没有中断合法操作。
  • 模糊案例(Ambiguous Cases): 成功将 全部 200 个模糊案例 升级,交由人在回路(human-in-the-loop)或高级仲裁处理。
  • 对抗韧性(Adversarial Resilience): 在攻击条件下,完整配置的 CONTINUITY 实现了 零有害外部副作用(zero harmful external effects)
  • Test Scale: Evaluated across 2,560 parameterized attack instances spanning 128 fault-domain classes.
  • Benign Performance: Successfully completed 100% of all 700 benign tasks without interrupting legitimate operations.
  • Ambiguous Cases: Successfully escalated all 200 ambiguous cases for human-in-the-loop or high-level arbitration.
  • Adversarial Resilience: The full CONTINUITY configuration committed zero harmful external effects under attack conditions.

4. 结论 (Conclusions)

这项研究表明,确保 LLM 智能体执行的安全需要的不仅仅是健全的单独安全控制。它需要明确的、可验证的契约,在从指令到效果的整个路径中维护安全保证。

The research demonstrates that securing LLM agent execution requires more than sound individual security controls. It demands explicit, verifiable contracts that preserve security guarantees across the complete instruction-to-effect path.


有关更多详细信息,请参考 arXiv 上的完整论文,或在 GitHub 上探索代码工件仓库。

For further details, reference the full paper on arXiv or explore the code artifact repository on GitHub.