ICANEWS

Engineering & Technology

SecureClaw: Dual-Boundary Security for LLM Agents Against External Actions and Plaintext Exposure

SecureClaw is presented as an architecture addressing security vulnerabilities in tool-using large language model (LLM) agents. These vulnerabilities are catego

Overview

SecureClaw is presented as an architecture addressing security vulnerabilities in tool-using large language model (LLM) agents. These vulnerabilities are categorized into two types: unauthorized external actions and the exposure of sensitive plaintext data within the agent's runtime environment prior to any final output checks. The architecture implements a dual-boundary defense mechanism to concurrently address both issues.

Research Context

Tool-using LLM agents are susceptible to specific security failures. Existing defense mechanisms commonly focus on protecting a single boundary, either the planner/runtime component or the action sink. This segmented protection means that such individual defenses do not comprehensively secure both surfaces simultaneously against the identified security threats. The problem originates from the need to manage sensitive information and external state changes within the LLM agent's operational environment.

Approach

SecureClaw employs a dual-boundary architecture. It positions authorization at the effect sink and implements plaintext confinement at the read boundary. In this design, sensitive read operations are routed through a trusted gateway. This gateway functions by replacing raw sensitive values with opaque handles. In the evaluated deployment configuration, this gateway also provides bounded summaries, which serve as an explicit declassification interface for sensitive data. Concurrently, write operations that modify external state are managed via a PREVIEW $\rightarrow$ COMMIT protocol. Under this protocol, only a trusted executor is empowered to commit the exact canonical request that has been authorized by policy. This architecture permits the runtime components of the LLM agent to plan operations using summaries and symbolic references, while preventing them from directly dereferencing sensitive secrets or initiating direct side effects on external systems.

The evaluation of SecureClaw involved comparative testing against existing defenses within a common harness across three distinct benchmarks:

  • AgentDojo
  • AgentLeak
  • Agent Security Bench (ASB)

Findings

During evaluation across the specified benchmarks, SecureClaw demonstrated specific performance characteristics:

  • On the Agent Security Bench (ASB), SecureClaw achieved a 0% attack success rate (ASR). This indicates its effectiveness in preventing successful attacks on this benchmark.
  • For AgentDojo, SecureClaw recorded a 0.64% ASR.
  • Pertaining to AgentLeak, the architecture showed a 3.23% overall leak on the attacked parity lane. This metric specifically measures leakage occurring both in final outputs and through internal relays.
  • The evaluation indicated that SecureClaw was the sole defense among those assessed in the common harness that simultaneously maintained usable task utility and achieved the anti-attack metrics cited above.

Specifically, the design allows the runtime to plan over summaries and symbolic references, while restricting direct dereferencing of secrets or execution of side effects.

Why This Matters

The observed dual-boundary protection provided by SecureClaw addresses security challenges faced by tool-using LLM agents, specifically unauthorized external actions and sensitive plaintext exposure within the runtime. The simultaneous achievement of low attack success rates across multiple benchmarks and retention of usable task utility suggests a potential advancement in securing these types of AI systems.

Key Findings