XiaoFeng Wang

Open-source uncensored large language models (ULLMs) are increasingly available in public LLM ecosystems. By removing or weakening safety guardrails, these models can comply with harmful requests, generate unsafe content, and be readily reused in downstream systems. While prior work has demonstrated the misuse potential of individual uncensored models, little has been done to systematically understand the ULLM ecosystem in real-world settings. In this paper, we present a systematic study of the open-source ULLM ecosystem. We analyzed 22,096 ULLMs collected from five major LLM hosting platforms and examined them across five interconnected components: artifacts, development, propagation, governance, and adoption. Our analysis shows that ULLMs are produced through four major methods, supported by a rich ecosystem of publicly available tools and uncensored datasets. We further found that ULLMs propagate widely across platforms, forming 4,457 cross-platform clusters comprising 11,094 models, with Hugging Face serving as the dominant upstream source. We uncovered systemic governance and security gaps, including inconsistent moderation labels applied to ULLMs across platforms, cross-platform propagation of ULLMs flagged with threat indicators, and frequent violations of base-model licensing constraints. ULLMs are also actively deployed in diverse real-world settings, including web applications, GitHub projects, underground forums, and model routing platforms, significantly lowering the barrier to harmful use. Our findings indicate that ULLMs are not isolated model artifacts but components of a broader ecosystem in which supply-chain dynamics, weak governance, and diverse downstream deployments jointly amplify risks.

Trusted Execution Environments (TEEs) have become a cornerstone of confidential computing, attracting significant attention from academia and industry. To support secure and scalable application deployment on confidential clouds, TEE containers (Tcons) have been introduced as middleware to shield applications from malicious operating systems and orchestration layers while preserving usability. In this paper, we present the first comprehensive analysis of Tcons, focusing on three critical layers: OS interfaces, encrypted I/O, and orchestration mechanisms. To enable systematic evaluation, we design TBouncer, an automated analyzer that precisely exercises and benchmarks Tcon isolation boundaries. Our study uncovers fundamental flaws in existing Tcons, leading to exploitable vulnerabilities such as code execution, denial-of-service, and information leakage. In total, we identify six attack vectors, twelve new bugs, and three CVEs. These findings provide new insights into the underestimated attack surface of Tcons and highlight key directions for building more secure and trustworthy container solutions.

Retrieval-Augmented Generation (RAG) systems are increasingly central to robust AI, enhancing large language model (LLM) faithfulness by incorporating external knowledge. However, our study unveils a critical, overlooked vulnerability: their profound susceptibility to subtle symbolic perturbations, particularly through near-imperceptible emotional icons (e.g., "(@_@)") that can catastrophically mislead retrieval, termed EmoRAG. We demonstrate that injecting a single emoticon into a query makes it nearly 100% likely to retrieve semantically unrelated texts, which contain a matching emoticon. Our extensive experiment across general question-answering and code domains, using a range of state-of-the-art retrievers and generators, reveals three key findings: (I) Single-Emoticon Disaster: Minimal emoticon injections cause maximal disruptions, with a single emoticon almost 100% dominating RAG output. (II) Positional Sensitivity: Placing an emoticon at the beginning of a query can cause severe perturbation, with F1-Scores exceeding 0.92 across all datasets. (III) Parameter-Scale Vulnerability: Counterintuitively, models with larger parameters exhibit greater vulnerability to the interference. We provide an in-depth analysis to uncover the underlying mechanisms of these phenomena. Furthermore, we raise a critical concern regarding the robustness assumption of current RAG systems, envisioning a threat scenario where an adversary exploits this vulnerability to manipulate the RAG system. We evaluate standard defenses and find them insufficient against EmoRAG. To address this, we propose targeted defenses, analyzing their strengths and limitations in mitigating emoticon-based perturbations. Finally, we outline future directions for building robust RAG systems.

Confidential computing (CC), designed for security-critical scenarios, uses remote attestation to guarantee code integrity on cloud servers. However, CC alone cannot provide assurance of high-level security properties (e.g., no data leak) on the code. In this paper, we introduce a novel framework, Agora, scrupulously designed to provide a trustworthy and open verification platform for CC. To prompt trustworthiness, we observe that certain verification tasks can be delegated to untrusted entities, while the corresponding (smaller) validators are securely housed within the trusted computing base (TCB). Moreover, through a novel blockchain-based bounty task manager, it also utilizes crowdsourcing to remove trust in complex theorem provers. These synergistic techniques successfully ameliorate the TCB size burden associated with two procedures: binary analysis and theorem proving. To prompt openness, Agora supports a versatile assertion language that allows verification of various security policies. Moreover, the design of Agora enables untrusted parties to participate in any complex processes out of Agora’s TCB. By implementing verification workflows for software-based fault isolation, information flow control, and side-channel mitigation policies, our evaluation demonstrates the efficacy of Agora.