Large language model (LLM) agents increasingly rely on external tools and retrieval systems to autonomously complete complex tasks. However, this design exposes agents to indirect prompt injection (IPI), where attacker-controlled context embedded in tool outputs or retrieved content silently steers agent actions away from user intent. Unlike prompt-based attacks, IPI unfolds over multi-turn trajectories, making malicious control difficult to disentangle from legitimate task execution. Existing inference-time defenses primarily rely on heuristic detection and conservative blocking of high-risk actions, which can prematurely terminate workflows or broadly suppress tool usage under ambiguous multi-turn scenarios. We propose AgentSentry, a novel inference-time detection and mitigation framework for tool-augmented LLM agents. To the best of our knowledge, AgentSentry is the first inference-time defense to model multi-turn IPI as a temporal causal takeover. It localizes takeover points via controlled counterfactual re-executions at tool-return boundaries and enables safe continuation through causally guided context purification that removes attack-induced deviations while preserving task-relevant evidence. We evaluate AgentSentry on the \textsc{AgentDojo} benchmark across four task suites, three IPI attack families, and multiple black-box LLMs. AgentSentry eliminates successful attacks and maintains strong utility under attack, achieving an average Utility Under Attack (UA) of 74.55 %, improving UA by 20.8 to 33.6 percentage points over the strongest baselines without degrading benign performance.

We present a novel approach for attacking black-box large language models (LLMs) by exploiting their ability to express confidence in natural language. Existing black-box attacks require either access to continuous model outputs like logits or confidence scores (which are rarely available in practice), or rely on proxy signals from other models. Instead, we demonstrate how to prompt LLMs to express their internal confidence in a way that is sufficiently calibrated to enable effective adversarial optimization. We apply our general method to three attack scenarios: adversarial examples for vision-LLMs, jailbreaks and prompt injections. Our attacks successfully generate malicious inputs against systems that only expose textual outputs, thereby dramatically expanding the attack surface for deployed LLMs. We further find that better and larger models exhibit superior calibration when expressing confidence, creating a concerning security paradox where model capability improvements directly enhance vulnerability. Our code is available at this [link](https://github.com/zj-jayzhang/black_box_llm_optimization).

The rise of manipulated media has made deepfakes a particularly insidious threat, involving various generative manipulations such as lip-sync modifications, face-swaps, and avatar-driven facial synthesis. Conventional detection methods, which predominantly depend on manually designed phoneme-viseme alignment thresholds, fundamental frame-level consistency checks, or a unimodal detection strategy, inadequately identify modern-day deepfakes generated by advanced generative models such as GANs, diffusion models, and neural rendering techniques. These advanced techniques generate nearly perfect individual frames yet inadvertently create minor temporal discrepancies frequently overlooked by traditional detectors. We present a novel multimodal audio-visual framework, Phoneme-Temporal and Identity-Dynamic Analysis(PIA), incorporating language, dynamic face motion, and facial identification cues to address these limitations. We utilize phoneme sequences, lip geometry data, and advanced facial identity embeddings. This integrated method significantly improves the detection of subtle deepfake alterations by identifying inconsistencies across multiple complementary modalities. Code is available at https://github.com/skrantidatta/PIA

Fine-tuning large language models (LLMs) with local data is a widely adopted approach for organizations seeking to adapt LLMs to their specific domains. Given the shared characteristics in data across different organizations, the idea of collaboratively fine-tuning an LLM using data from multiple sources presents an appealing opportunity. However, organizations are often reluctant to share local data, making centralized fine-tuning impractical. Federated learning (FL), a privacy-preserving framework, enables clients to retain local data while sharing only model parameters for collaborative training, offering a potential solution. While fine-tuning LLMs on centralized datasets risks data leakage through next-token prediction, the iterative aggregation process in FL results in a global model that encapsulates generalized knowledge, which some believe protects client privacy. In this paper, however, we present contradictory findings through extensive experiments. We show that attackers can still extract training data from the global model, even using straightforward generation methods, with leakage increasing as the model size grows. Moreover, we introduce an enhanced attack strategy tailored to FL, which tracks global model updates during training to intensify privacy leakage. To mitigate these risks, we evaluate privacy-preserving techniques in FL, including differential privacy, regularization-constrained updates and adopting LLMs with safety alignment. Our results provide valuable insights and practical guidelines for reducing privacy risks when training LLMs with FL.

Adversarial attacks from generative models often produce low-quality images and require substantial computational resources. Diffusion models, though capable of high-quality generation, typically need hundreds of sampling steps for adversarial generation. This paper introduces TAIGen, a training-free black-box method for efficient adversarial image generation. TAIGen produces adversarial examples using only 3-20 sampling steps from unconditional diffusion models. Our key finding is that perturbations injected during the mixing step interval achieve comparable attack effectiveness without processing all timesteps. We develop a selective RGB channel strategy that applies attention maps to the red channel while using GradCAM-guided perturbations on green and blue channels. This design preserves image structure while maximizing misclassification in target models. TAIGen maintains visual quality with PSNR above 30 dB across all tested datasets. On ImageNet with VGGNet as source, TAIGen achieves 70.6% success against ResNet, 80.8% against MNASNet, and 97.8% against ShuffleNet. The method generates adversarial examples 10x faster than existing diffusion-based attacks. Our method achieves the lowest robust accuracy, indicating it is the most impactful attack as the defense mechanism is least successful in purifying the images generated by TAIGen.