CAPMAS: Capability-Based Delegation of Privileges in Multi-Agent Systems
CAPMAS通过结合对比学习语义范围管道和Macaroon令牌，解决了多代理系统中特权安全高效委托的问题，实现了更快的委托操作与更低的延迟及带宽使用。
0 citationsRead paper
Selected work
Representative Papers
Aether: Network Validation Using Agentic AI and Digital Twin
Network change validation has long relied on manual processes, which are inefficient and error-prone, while existing approaches struggle to handle the complexity of continuous changes in real-world production environments. This work proposes the first end-to-end automated validation framework that integrates multi-agent generative AI with a unified network digital twin. The framework employs five specialized agents that collaboratively orchestrate the entire pipeline—from intent parsing to test validation—by tightly coupling modeling, simulation, and agent coordination. Evaluated on both synthetic scenarios and a real ISP network, the system achieves 100% error detection accuracy and 92–96% diagnostic coverage, completing validation in just 6–7 minutes, substantially outperforming conventional methods.
0 citationsRead paper
Trojan horse hunt in deep forecasting models: Insights from the European Space Agency competition
This work addresses the vulnerability of deep time series forecasting models to Trojan backdoor attacks in safety-critical applications such as spacecraft telemetry, where specific trigger patterns can maliciously manipulate predictions. To tackle this emerging threat, the study pioneers the introduction of backdoor detection into the time series domain by organizing an international data science competition focused on temporal models. The effort establishes novel benchmark tasks, evaluation protocols, and a public dataset, integrating techniques from adversarial example analysis, model reverse engineering, and time series anomaly detection to investigate trigger localization and backdoor verification. The competition attracted over 200 participating teams, yielding diverse and effective detection strategies, key research insights, and a fully open-sourced repository of materials, thereby laying a foundation for developing secure and reliable deep time series forecasting systems.
0 citationsRead paper
Design Patterns for Securing LLM Agents against Prompt Injections
Large language model (LLM) agents are vulnerable to prompt injection attacks, particularly during tool invocation and sensitive data handling, posing critical security risks. Method: This paper introduces the first provably secure design pattern system specifically for prompt injection defense. We formalize security requirements, conduct rigorous threat modeling, and abstract seven core defense patterns—covering input sanitization, context isolation, execution sandboxing, and other critical mitigation pathways—while formally quantifying the trade-offs between security guarantees and functional utility for each pattern. Contribution/Results: All patterns are validated on real-world, industrial-grade agent architectures. Empirical evaluation demonstrates that they effectively block high-severity prompt injection attacks and significantly enhance system robustness without compromising operational functionality.
0 citationsRead paper
Recent publications
Latest Papers
CAPMAS: Capability-Based Delegation of Privileges in Multi-Agent Systems
CAPMAS通过结合对比学习语义范围管道和Macaroon令牌，解决了多代理系统中特权安全高效委托的问题，实现了更快的委托操作与更低的延迟及带宽使用。
0 citationsRead paper
Aether: Network Validation Using Agentic AI and Digital Twin
Network change validation has long relied on manual processes, which are inefficient and error-prone, while existing approaches struggle to handle the complexity of continuous changes in real-world production environments. This work proposes the first end-to-end automated validation framework that integrates multi-agent generative AI with a unified network digital twin. The framework employs five specialized agents that collaboratively orchestrate the entire pipeline—from intent parsing to test validation—by tightly coupling modeling, simulation, and agent coordination. Evaluated on both synthetic scenarios and a real ISP network, the system achieves 100% error detection accuracy and 92–96% diagnostic coverage, completing validation in just 6–7 minutes, substantially outperforming conventional methods.
0 citationsRead paper
Trojan horse hunt in deep forecasting models: Insights from the European Space Agency competition
This work addresses the vulnerability of deep time series forecasting models to Trojan backdoor attacks in safety-critical applications such as spacecraft telemetry, where specific trigger patterns can maliciously manipulate predictions. To tackle this emerging threat, the study pioneers the introduction of backdoor detection into the time series domain by organizing an international data science competition focused on temporal models. The effort establishes novel benchmark tasks, evaluation protocols, and a public dataset, integrating techniques from adversarial example analysis, model reverse engineering, and time series anomaly detection to investigate trigger localization and backdoor verification. The competition attracted over 200 participating teams, yielding diverse and effective detection strategies, key research insights, and a fully open-sourced repository of materials, thereby laying a foundation for developing secure and reliable deep time series forecasting systems.
0 citationsRead paper
Design Patterns for Securing LLM Agents against Prompt Injections
Large language model (LLM) agents are vulnerable to prompt injection attacks, particularly during tool invocation and sensitive data handling, posing critical security risks. Method: This paper introduces the first provably secure design pattern system specifically for prompt injection defense. We formalize security requirements, conduct rigorous threat modeling, and abstract seven core defense patterns—covering input sanitization, context isolation, execution sandboxing, and other critical mitigation pathways—while formally quantifying the trade-offs between security guarantees and functional utility for each pattern. Contribution/Results: All patterns are validated on real-world, industrial-grade agent architectures. Empirical evaluation demonstrates that they effectively block high-severity prompt injection attacks and significantly enhance system robustness without compromising operational functionality.
0 citationsRead paper