A new type of attack, Micro-Collaborative Poisoning, is introduced that targets Retrieval-Augmented Generation (RAG) systems. This attack involves dividing a false target claim across multiple documents, making it difficult to detect. The attack is evaluated across various RAG configurations, and the results show that it can be effective even with diverse and strong retrievers. This highlights the need for improved defenses against such attacks.

Researchers introduced a new framework for detecting zero-day jamming attacks, which combines graph attention networks and Dirichlet process clustering. They also proposed an inference-driven reinforcement learning-based jammer that can evade detection. These advancements are relevant to the AI agent security landscape, particularly in the context of Model Context Protocol (MCP) security.

This research paper introduces a new approach to adapting deep learning models for underwater synthetic aperture sonar Automatic Target Recognition (ATR) using a three-stage framework that leverages Low-Rank Adaptation (LoRA) and Supervised Contrastive Learning (SupCon). The method improves performance by 379% on a mission-level evaluation, with a significant reduction in training data required.

Batch normalization amplifies memorization of outliers and increases susceptibility to membership inference attacks, posing a privacy risk to models using this technique. Researchers conducted an empirical study on multiple datasets and architectures, observing a higher memorization of outliers and increased susceptibility to attacks.

This research explores how the choice of p in lp norms affects the sparsity and smoothness of adversarial attacks against deep neural networks. The study finds that the optimal p value depends on the task and that lp norms with p in [1.3, 1.5] yield the best trade-off between sparse and smooth attacks.

Researchers proposed Inverse Knowledge Distillation (IKD), a mechanism to enhance adversarial transferability by maximizing the prediction-distribution discrepancy between benign and adversarial samples. This method uses a soft-label objective to enrich attacks with surrogate directions, and its analysis derives a lower bound on Fisher-subspace overlap between surrogate and target models.

Researchers propose AttnPrint, a white-box provenance method for multimodal large language models, and DistillTrace, a black-box auditing tool to detect model infringement and distillation. These solutions aim to safeguard model ownership and prevent illicit deployment and unauthorized distillation of MLLMs.

Researchers propose CoReLoop, a method for refining SSL-based audio deepfake detectors without additional training data or parameters. CoReLoop adapts recurrent inputs to the frozen encoder and controls state updates to improve detection performance. This advancement could be relevant to developers working on audio deepfake detection and mitigation.

Researchers developed FARSIGHT, a framework to evaluate the robustness and security of financial LLM agents against market turbulence and adversarial attacks. They applied it to 15 academic schemes, finding most failed to account for robustness and security, leaving them vulnerable to market crashes.

Researchers identify a readout gap in vision-language models for harmful meme detection, where models miss internal evidence or struggle to route represented evidence to the output. They propose using sparse autoencoders and role-conditioned probes to improve performance, achieving significant gains in macro-F1 scores. This affects AI agents by highlighting a common bottleneck in harmful content classification and suggesting techniques to address it.

Research proposes a framework, Know Your Agent (KYA), for reconnaissance-driven pentesting of AI agents to identify weaknesses and craft stronger attacks. The framework automates black-box testing by probing agents, building target profiles, and creating more effective attacks. This matters to AI agent operators as it highlights the need for similar reconnaissance techniques to protect against indirect prompt injection attacks.

Researchers developed a method called SALVE to detect subliminal learning effects in AI models by using text optimization to recover legible prompts from a distillation dataset. This method can identify traits from the teacher model not encoded in the dataset, and can be used to detect subliminal learning effects in various settings.

Researchers propose two universal tool-based defenses (Attacker Tool Filtering and Normal Tool Recalling) and prompt-based defenses (Chain-of-Thought prompting and self-reflection techniques) to mitigate adversarial attacks on tool-integrated LLM agents. Experimental results show significant reduction in Attack Success Rates while preserving or improving task success rates across various LLMs. The code is available on GitHub.