The Framing Gap: Indirect Prompt-Injection Exfiltration Defeats Surface-Level Defenses in Tool-Using Agents
研究解决了工具使用型LLM代理在读取攻击者控制的内容时可能泄露秘密的问题，通过分析不同防御方法的有效性，发现限制目的地或隔离能力是有效的解决策略。
0 citationsRead paper
Selected work
Representative Papers
Logical Neural Belief Propagation for Linear-Complexity Decoding of Surface Codes
本文提出逻辑神经置信传播(L-NBP)解码器，通过将解码目标从物理层转向逻辑层，并结合神经网络权重训练，实现线性复杂度下的高精度量子纠错。
0 citationsRead paper
Proof-of-Execution Memory: Defending LLM Agents Against Forged-Reasoning Attacks by Verifying What Actually Happened
This study addresses the vulnerability of LLM agents to reasoning forgery attacks and the limitations of existing content-based defenses. We propose Proof-of-Execution Memory (PoEM), a novel cryptographic verification mechanism that replaces content detection with HMAC-chained ledgers and trusted-layer write controls. This approach fundamentally prevents memory tampering and ensures immunity to text-paraphrasing attacks. Experiments across three models and scenarios demonstrate that PoEM reduces attack success rates to 0% with zero false positives, significantly outperforming SENTINEL. With microsecond-level verification overhead, PoEM effectively resolves the capability paradox wherein more powerful models exhibit greater susceptibility to attacks, thereby providing mechanism-level security guarantees for autonomous agents.
0 citationsRead paper
HBRB-BoW: A Retrained Bag-of-Words Vocabulary for ORB-SLAM via Hierarchical BRB-KMeans
This work addresses the degradation in clustering accuracy and loss of discriminative power in the binary Bag-of-Words (BoW) model used in traditional ORB-SLAM, which stems from premature binarization that compromises both the distinctiveness and structural integrity of visual vocabulary, thereby impairing loop closure detection and relocalization performance. To mitigate this issue, the authors propose the HBRB-BoW algorithm, which integrates a global real-valued information flow throughout the hierarchical clustering process and defers final binarization exclusively to the leaf nodes. The approach further incorporates an enhanced hierarchical BRB-KMeans clustering scheme and an optimized ORB descriptor strategy. This design effectively alleviates the loss of fine-grained feature details, substantially improving the representational capacity of the BoW model under challenging environmental conditions and consequently enhancing the robustness of loop closure detection and relocalization in ORB-SLAM.
0 citationsRead paper
TT-FSI: Scalable Faithful Shapley Interactions via Tensor-Train
This work addresses the scalability limitations of the Faithful Shapley Interaction (FSI) index, which suffers from exponential computational and memory costs in high-dimensional settings. We reveal for the first time that the FSI linear operator admits a low-rank matrix product operator (MPO) structure with a Tensor-Train rank of only O(ℓd). Leveraging this insight, we design an efficient sweeping algorithm that achieves exponential reductions in both time and space complexity. Empirical evaluations across six datasets with coalition sizes ranging from d=8 to d=20 demonstrate up to 280× speedup and 290× memory reduction compared to baseline methods, and an 85× improvement over SHAP-IQ. Our approach successfully scales FSI to coalitions of up to one million players (d=20), marking a significant advance in practical applicability.
0 citationsRead paper
Recent publications
Latest Papers
The Framing Gap: Indirect Prompt-Injection Exfiltration Defeats Surface-Level Defenses in Tool-Using Agents
研究解决了工具使用型LLM代理在读取攻击者控制的内容时可能泄露秘密的问题，通过分析不同防御方法的有效性，发现限制目的地或隔离能力是有效的解决策略。
0 citationsRead paper
Logical Neural Belief Propagation for Linear-Complexity Decoding of Surface Codes
本文提出逻辑神经置信传播(L-NBP)解码器，通过将解码目标从物理层转向逻辑层，并结合神经网络权重训练，实现线性复杂度下的高精度量子纠错。
0 citationsRead paper
Proof-of-Execution Memory: Defending LLM Agents Against Forged-Reasoning Attacks by Verifying What Actually Happened
This study addresses the vulnerability of LLM agents to reasoning forgery attacks and the limitations of existing content-based defenses. We propose Proof-of-Execution Memory (PoEM), a novel cryptographic verification mechanism that replaces content detection with HMAC-chained ledgers and trusted-layer write controls. This approach fundamentally prevents memory tampering and ensures immunity to text-paraphrasing attacks. Experiments across three models and scenarios demonstrate that PoEM reduces attack success rates to 0% with zero false positives, significantly outperforming SENTINEL. With microsecond-level verification overhead, PoEM effectively resolves the capability paradox wherein more powerful models exhibit greater susceptibility to attacks, thereby providing mechanism-level security guarantees for autonomous agents.
0 citationsRead paper
HBRB-BoW: A Retrained Bag-of-Words Vocabulary for ORB-SLAM via Hierarchical BRB-KMeans
This work addresses the degradation in clustering accuracy and loss of discriminative power in the binary Bag-of-Words (BoW) model used in traditional ORB-SLAM, which stems from premature binarization that compromises both the distinctiveness and structural integrity of visual vocabulary, thereby impairing loop closure detection and relocalization performance. To mitigate this issue, the authors propose the HBRB-BoW algorithm, which integrates a global real-valued information flow throughout the hierarchical clustering process and defers final binarization exclusively to the leaf nodes. The approach further incorporates an enhanced hierarchical BRB-KMeans clustering scheme and an optimized ORB descriptor strategy. This design effectively alleviates the loss of fine-grained feature details, substantially improving the representational capacity of the BoW model under challenging environmental conditions and consequently enhancing the robustness of loop closure detection and relocalization in ORB-SLAM.
0 citationsRead paper
TT-FSI: Scalable Faithful Shapley Interactions via Tensor-Train
This work addresses the scalability limitations of the Faithful Shapley Interaction (FSI) index, which suffers from exponential computational and memory costs in high-dimensional settings. We reveal for the first time that the FSI linear operator admits a low-rank matrix product operator (MPO) structure with a Tensor-Train rank of only O(ℓd). Leveraging this insight, we design an efficient sweeping algorithm that achieves exponential reductions in both time and space complexity. Empirical evaluations across six datasets with coalition sizes ranging from d=8 to d=20 demonstrate up to 280× speedup and 290× memory reduction compared to baseline methods, and an 85× improvement over SHAP-IQ. Our approach successfully scales FSI to coalitions of up to one million players (d=20), marking a significant advance in practical applicability.
0 citationsRead paper