Benign fine-tuning severely weakens the safety alignment of large language models (LLMs), so we study why refusal behavior is so fragile. While prior work often attributes this failure to gradient conflict, we propose a fundamentally different Fisher-geometric explanation: safety Fisher is low-rank, and alignment makes the safety geometry flatter while preserving an output-routing pathway. After 100 benign fine-tuning examples, this pathway is selectively re-sharpened in output-side MLP modules, explaining the asymmetric fragility: safety can collapse to high attack success rates, while general utility degrades mildly.

The routing view also explains why few safety examples can restore refusal behavior, indicating that internal safety-relevant representations are preserved. Finally, we show that LoRA and ASAM mitigate early collapse by suppressing output-side sharpness, but their protection weakens at larger fine-tuning scales. Overall, safety failure is best understood as a disruption of a low-rank output-routing mechanism.

Large language models (LLMs) often memorize personally identifiable information (PII) during training, raising serious privacy concerns. While machine unlearning has emerged as a promising mitigation, its effectiveness against privacy attacks remains unclear. We introduce PrivUn, an evaluation framework for privacy unlearning under recovery attacks, paired with three measurement tools: forgetting scores, association metrics, and layer-wise analysis of forgetting depth. This framework studies not only whether a model forgets, but also how forgetting propagates and where it fails.

Using PrivUn, we make two findings. (i) The PII ripple effect is gradient-driven, not semantic. Graph-based proximity is uncorrelated with PII forgetting, while gradient-based association is more strongly correlated. (ii) Label-based unlearning is typically shallow, while PII is stored across deeper layers. Layer-wise evidence suggests that label-based methods mainly change the final layers, explaining why they reduce direct retrieval but remain vulnerable to stronger recovery attacks.

As text-to-music models gain widespread adoption, the prompts used to guide these systems have become valuable intellectual property. This shift has given rise to a new form of attack: prompt stealing, aiming to reconstruct the high-value prompts that guide the music generation. However, unlike prior work in text and image generation, prompt stealing in text-to-music systems faces unique challenges due to the entangled and diffuse nature of semantic representations in audio, which complicates the decoupling of specific textual tokens from acoustic outputs. To address these challenges, we present AudioStealer, the first targeted study of prompt inversion in the audio domain. AudioStealer operates via a two-stage black-box attack framework: first, a heuristic search guided by audio-language embeddings identifies initial candidates; then, these candidates are refined using a game-theoretic strategy based on Shapley value estimation to attribute precise semantic contributions. Our method requires no direct access to the target model and relies solely on a shadow model, making it broadly applicable. Through extensive experiments, we demonstrate that AudioStealer recovers prompts with high textual consistency to the ground truth, while the regenerated audio maintains strong perceptual similarity to the target recordings. These results expose critical vulnerabilities in the text-to-audio market ecosystem and underscore the urgent need for intellectual property protections in generative audio technologies.

Cloud-hosted Large Language Models (LLMs) offer unmatched reasoning capabilities and dynamic knowledge, yet submitting raw queries to these external services risks exposing sensitive user intent. Conversely, relying exclusively on trusted local models preserves privacy but often compromises answer quality due to limited parameter scale and knowledge. To resolve this dilemma, we propose Game-theoretic Trustworthy Knowledge Acquisition (GTKA), a framework that formulates the trade-off between knowledge utility and privacy as a strategic game. GTKA consists of three components: (i) a privacy-aware sub-query generator that decomposes sensitive intent into generalized, low-risk fragments; (ii) an adversarial reconstruction attacker that attempts to infer the original query from these fragments, providing adaptive leakage signals; and (iii) a trusted local integrator that synthesizes external responses within a secure boundary. By training the generator and attacker in an alternating adversarial manner, GTKA optimizes the sub-query generation policy to maximize knowledge acquisition accuracy while minimizing the reconstructability of the original sensitive intent. To validate our approach, we construct two sensitive-domain benchmarks in the biomedical and legal fields. Extensive experiments demonstrate that GTKA significantly reduces intent leakage compared to state-of-the-art baselines while maintaining high-fidelity answer quality.

Reasoning-enhanced large language models rely on intermediate reasoning signals to solve complex, multi-step tasks, making reasoning behavior a valuable form of intellectual property. Meanwhile, knowledge distillation enables an adversary to replicate this behavior in a realistic black-box setting by repeatedly querying a deployed model on a target domain and training a local student to imitate its outputs, including reasoning traces. Existing LLM watermarks primarily operate on surface text and decoding-time token biases, and thus fail to provide reliable attribution of reasoning behavior once it is transferred through knowledge distillation. ReasMark entangles the watermark with the target-domain input distribution by selecting watermark tokens from high-frequency prompts, so distillation queries naturally activate it. It then embeds the watermark by score-conditioned losses that create a detectable reasoning-length gap for black-box verification. Comprehensive experiments across multiple LLMs, datasets, and distillation settings demonstrate that ReasMark consistently outperforms existing baselines while preserving task utility.

Safety alignment of large language models currently faces a central challenge: existing alignment techniques often prioritize mitigating responses to harmful prompts at the expense of overcautious behavior, leading models to incorrectly refuse benign requests. A key goal of safe alignment is therefore to improve safety while simultaneously minimizing false refusals. In this work, we introduce Energy Landscape Steering (ELS), a novel, fine-tuning free framework designed to resolve this challenge through dynamic, inference-time intervention. We trained a lightweight, external Energy-Based Model (EBM) to assign high energy to undesirable (false refusal or jailbreak) states and low energy to desirable (helpful response or safe reject) ones. During inference, the EBM maps the LLM’s internal activations to an energy landscape, and we use the gradient of the energy function to steer the hidden states toward low-energy regions in real time. This dynamically guides the model toward desirable behavior without modifying its parameters. By decoupling behavioral control from the model’s core knowledge, ELS provides a flexible and computationally efficient solution. Extensive experiments across diverse models demonstrate its effectiveness: raising compliance on the ORB-H benchmark from 57.3% to 82.6% while maintaining the baseline safety performance. Our work establishes a promising paradigm for building LLMs that simultaneously achieve high safety and low false refusal rates.