Yuansen Zhang

Large Language Models (LLMs) demonstrate strong capabilities in solving complex tasks when integrated with external tools. The Model Context Protocol (MCP) has become a standard interface for enabling such tool-based interactions. However, these interactions introduce substantial security concerns, particularly when the MCP server is compromised or untrustworthy. While prior benchmarks primarily focus on prompt injection attacks or analyze the vulnerabilities of LLM-MCP interaction trajectories, limited attention has been given to the underlying system logs associated with malicious MCP servers. To address this gap, we present the first synthetic benchmark for evaluating LLMs' ability to identify security risks from system logs. We define nine categories of MCP server risks and generate 1,800 synthetic system logs using ten state-of-the-art LLMs. These logs are embedded in the return values of 243 curated MCP servers, yielding a dataset of 2,421 chat histories for training and 471 queries for evaluation. Our pilot experiments reveal that smaller models often fail to detect risky system logs, leading to high false negatives. While models trained with supervised fine-tuning (SFT) tend to over-flag benign logs, resulting in elevated false positives, Reinforcement Learning with Verifiable Reward (RLVR) offers a better precision-recall balance. In particular, after training with Group Relative Policy Optimization (GRPO), Llama3.1-8B-Instruct achieves 83 percent accuracy, surpassing the best-performing large remote model by 9 percentage points. Fine-grained, per-category analysis further underscores the effectiveness of reinforcement learning in enhancing LLM safety within the MCP framework. Code and data are available at https://github.com/PorUna-byte/MCP-RiskCue.

Large language models often face a three-way trade-off among detection accuracy, inference latency, and deployment cost when used in real-world safety-sensitive applications. This paper introduces Prefix Probing, a black-box harmful content detection method that compares the conditional log-probabilities of "agreement/execution" versus "refusal/safety" opening prefixes and leverages prefix caching to reduce detection overhead to near first-token latency. During inference, the method requires only a single log-probability computation over the probe prefixes to produce a harmfulness score and apply a threshold, without invoking any additional models or multi-stage inference. To further enhance the discriminative power of the prefixes, we design an efficient prefix construction algorithm that automatically discovers highly informative prefixes, substantially improving detection performance. Extensive experiments demonstrate that Prefix Probing achieves detection effectiveness comparable to mainstream external safety models while incurring only minimal computational cost and requiring no extra model deployment, highlighting its strong practicality and efficiency.