Safety alignment in large audio-language models (LALMs) is typically evaluated through single-turn interactions with direct prompts. In real-world deployments, however, these models are often integrated into agent frameworks that introduce developer-controlled system prompts, tool-calling loops, and intermediate encoding steps between user speech and model input. This raises a practical question: can a text-based jailbreak remain effective after being converted to speech by an off-the-shelf text-to-speech (TTS) system and processed through an agent? We investigate this question using 270 prohibited questions spanning nine policy categories, an open-weight TTS model with 82 million parameters, and a minimal LangChain question-answering agent whose system prompt combines a persona with a fuzzing-derived template. Across three audio-capable models, we observe attack success rates (ASRs) ranging from 57.0% to 64.1%. To assess the impact of synthesis errors, we use automatic speech recognition to verify the generated audio and find that 87.4% of clips are transcribed exactly. Our analysis indicates that faithfully rendered speech alone achieves an ASR above 50% on every model tested, while manual evaluation of a 50-response sample agrees with the automatic judge in 92% of cases. These findings suggest that the agent framework does not compensate for the loss of safety robustness introduced by the text-to-audio transition. Even a readily available TTS pipeline requiring only minutes of computation can therefore provide a practical attack surface for transferring text-based jailbreaks to audio-agent settings.

KV caching significantly improves the efficiency of Large Language Model (LLM) inference by storing attention states from previously processed tokens, enabling faster generation of subsequent tokens. However, as sequence length increases, the KV cache quickly becomes a major memory bottleneck. To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM’s PagedAttention. Unlike existing approaches that rely on attention-based token importance or evict tokens across different vLLM pages, PagedEviction introduces an efficient block-wise eviction algorithm tailored for paged memory layouts. Our method integrates seamlessly with PagedAttention without requiring any modifications to its CUDA attention kernels. We evaluate PagedEviction across Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, and Llama-3.2-3B-Instruct models on the LongBench benchmark suite, demonstrating improved memory usage with better accuracy than baselines on long context tasks.

Artificial Intelligence (AI) and Large Language Models (LLMs) are increasingly integrated into high-stakes applications, yet their susceptibility to adversarial prompts poses significant security risks. In this work, we introduce Mātṛkā, a framework for systematically evaluating jailbreak vulnerabilities in open-source multilingual LLMs. Using the open-source dataset across nine sensitive categories, we constructed adversarial prompt sets that combine translation, mixed-language encoding, homoglyph signatures, numeric enforcement, and structural variations. Experiments were conducted on state-of-the-art open-source models from Llama, Qwen, GPT-OSS, Mistral, and Gemma families. Our findings highlight transferability of jailbreaks across multiple languages with varying success rates depending on attack design. We provide empirical insights, a novel taxonomy of multilingual jailbreak strategies, and recommendations for enhancing robustness in safety-critical environments.