The Intelligence Advanced Research Projects Activity (IARPA) launched the TrojAI program to confront an emerging vulnerability in modern artificial intelligence: the threat of AI Trojans. These AI trojans are malicious, hidden backdoors intentionally embedded within an AI model that can cause a system to fail in unexpected ways, or allow a malicious actor to hijack the AI model at will. This multi-year initiative helped to map out the complex nature of the threat, pioneered foundational detection methods, and identified unsolved challenges that require ongoing attention by the burgeoning AI security field. This report synthesizes the program's key findings, including methodologies for detection through weight analysis and trigger inversion, as well as approaches for mitigating Trojan risks in deployed models. Comprehensive test and evaluation results highlight detector performance, sensitivity, and the prevalence of "natural" Trojans. The report concludes with lessons learned and recommendations for advancing AI security research.

Large Language Models (LLMs) are increasingly employed in various question-answering tasks. However, recent studies showcase that LLMs are susceptible to persuasion and could adopt counterfactual beliefs. We present a systematic evaluation of LLM susceptibility to persuasion under the Source--Message--Channel--Receiver (SMCR) communication framework. Across five mainstream Large Language Models (LLMs) and three domains (factual knowledge, medical QA, and social bias), we analyze how different persuasive strategies influence belief stability over multiple interaction turns. We further examine whether meta-cognition prompting (i.e., eliciting self-reported confidence) affects resistance to persuasion. Results show that the smallest model (Llama 3.2-3B) exhibits extreme compliance, with 82.5% of belief changes occurring at the first persuasive turn (average end turn of 1.1--1.4). Contrary to expectations, meta-cognition prompting increases vulnerability by accelerating belief erosion rather than enhancing robustness. Finally, we evaluate adversarial fine-tuning as a defense. While GPT-4o-mini achieves near-complete robustness (98.6%) and Mistral~7B improves substantially (35.7% $\rightarrow$ 79.3%), Llama models remain highly susceptible (<14%) even when fine-tuned on their own failure cases. Together, these findings highlight substantial model-dependent limits of current robustness interventions and offer guidance for developing more trustworthy LLMs.

Recommender systems play a central role in digital platforms by providing personalized content. They often use methods such as collaborative filtering and machine learning to accurately predict user preferences. Although these systems offer substantial benefits, they are vulnerable to security and privacy threats, especially data poisoning attacks. By inserting misleading data, attackers can manipulate recommendations for purposes ranging from boosting product visibility to shaping public opinion. Despite these risks, concerns are often downplayed because such attacks typically require controlling at least 1% of the platform's user base, a difficult task on large platforms. We tackle this issue by introducing the IndirectAD attack, inspired by Trojan attacks on machine learning. IndirectAD reduces the need for a high poisoning ratio through a trigger item that is easier to recommend to the target users. Rather than directly promoting a target item that does not match a user's interests, IndirectAD first promotes the trigger item, then transfers that advantage to the target item by creating co-occurrence data between them. This indirect strategy delivers a stronger promotion effect while using fewer controlled user accounts. Our extensive experiments on multiple datasets and recommender systems show that IndirectAD can cause noticeable impact with only 0.05% of the platform's user base. Even in large-scale settings, IndirectAD remains effective, highlighting a more serious and realistic threat to today's recommender systems.

With the rapid development of large language models, the potential threat of their malicious use, particularly in generating phishing content, is becoming increasingly prevalent. Leveraging the capabilities of LLMs, malicious users can synthesize phishing emails that are free from spelling mistakes and other easily detectable features. Furthermore, such models can generate topic-specific phishing messages, tailoring content to the target domain and increasing the likelihood of success. Detecting such content remains a significant challenge, as LLM-generated phishing emails often lack clear or distinguishable linguistic features. As a result, most existing semantic-level detection approaches struggle to identify them reliably. While certain LLM-based detection methods have shown promise, they suffer from high computational costs and are constrained by the performance of the underlying language model, making them impractical for large-scale deployment. In this work, we aim to address this issue. We propose Paladin, which embeds trigger-tag associations into vanilla LLM using various insertion strategies, creating them into instrumented LLMs. When an instrumented LLM generates content related to phishing, it will automatically include detectable tags, enabling easier identification. Based on the design on implicit and explicit triggers and tags, we consider four distinct scenarios in our work. We evaluate our method from three key perspectives: stealthiness, effectiveness, and robustness, and compare it with existing baseline methods. Experimental results show that our method outperforms the baselines, achieving over 90% detection accuracy across all scenarios.

Global KV-cache sharing is an effective optimization for accelerating large language model (LLM) inference, yet it introduces an API-visible timing side channel that lets adversaries infer sensitive user inputs from shared entries, leading to cross-tenant privacy risks. To address this problem, we introduce SafeKV (Secure and Flexible KV-cache Sharing), a system-level co-design of privacy enforcement and KV-cache management. SafeKV integrates lightweight detection and isolation directly into the serving runtime to eliminate cross-tenant reuse of sensitive KV-cache blocks under our threat model, while recovering most of the performance benefits of global sharing. Our key contributions are: (1) a three-tier asynchronous detection pipeline that decouples privacy classification from inference and supports streaming workloads, (2) a unified radix-tree-based memory manager with path compression and sensitivity-aware eviction for scalable selective isolation, and (3) an RDR-guided (Reuse Diversity Ratio) runtime safeguard that detects and bounds residual leakage. On large LLM backends, SafeKV reduces the time-to-first-token (TTFT) overhead compared to full isolation by up to 40.58% and raises throughput by up to 2.66x. Overall, SafeKV restores the efficiency of KV reuse while enforcing strong, practical privacy for multi-tenant LLM inference.

In recent years, tremendous success has been witnessed in Retrieval-Augmented Generation (RAG), widely used to enhance Large Language Models (LLMs) in domain-specific, knowledge-intensive, and privacy-sensitive tasks. However, attackers may steal those valuable RAGs and deploy or commercialize them, making it essential to detect Intellectual Property (IP) infringement. Most existing ownership protection solutions, such as watermarks, are designed for relational databases and texts. They cannot be directly applied to RAGs because relational database watermarks require white-box access to detect IP infringement, which is unrealistic for the knowledge base in RAGs. Meanwhile, post-processing by the adversary's deployed LLMs typically destructs text watermark information. To address those problems, we propose a novel black-box "knowledge watermark" approach, named RAG-WM, to detect IP infringement of RAGs. RAG-WM uses a multi-LLM interaction framework, comprising a Watermark Generator, Shadow LLM & RAG, and Watermark Discriminator, to create watermark texts based on watermark entity-relationship tuples and inject them into the target RAG. We evaluate RAG-WM across three domain-specific and two privacy-sensitive tasks on four benchmark LLMs. Experimental results show that RAG-WM effectively detects the stolen RAGs in various deployed LLMs. Furthermore, RAG-WM is robust against paraphrasing, unrelated content removal, knowledge insertion, and knowledge expansion attacks. Lastly, RAG-WM can also evade watermark detection approaches, highlighting its promising application in detecting IP infringement of RAG systems.