MuPPET introduces a benchmark for contextual privacy in multi-party conversations. Experiments reveal models leak significantly more private information in group settings than in one-to-one interactions, with smaller open-weights models being especially vulnerable. Existing privacy defenses provide only partial protection and fail to address the core issue of party tracking.
We propose Uncertainty-Based Decontamination (UBD), a method that uses deep ensembles to estimate per-sample memorization in contaminated models without needing an uncontaminated model. UBD constructs a debiased target distribution from ensemble uncertainty to correct output distributions, achieving significantly better alignment with uncontaminated models compared to baselines, while maintaining performance on clean data.
TF-RefusalBench is a multilingual benchmark derived from Swiss Supreme Court rulings, containing 5,200 prompts in French, German, Italian, and English. It reveals that over-alignment in LLMs is influenced by model and language factors, and that refusals impact task faithfulness beyond simple refusal rates. Abliteration of refusal directives reduces over-alignment with minimal performance loss in criminal law tasks.
A study of 1,174 Reddit users reveals four distinct self-stigma personas. LLMs trained to recognize these personas outperform generic models in targeted responses, though clinical experts prefer generic empathy over persona-matched support. The research highlights a tension between tailored empathy and holistic user preference in stigma-related AI interventions.
Open language models show evaluation awareness is not a unified trait. Eight experiments across 37 models reveal detection, safety behavior shifts, and representation stability vary independently, with only weak correlations between them. This undermines the idea of a single awareness score as a reliable indicator of deployment safety, highlighting the 'benchmark illusion'.
No large language models reliably detect when their responses were influenced by adversarial prefill attacks. Introspective signals are strongest in safety-related reasoning, but are probe-dependent and can be amplified by LoRA fine-tuning, which paradoxically increases attack success rates.
The EU AI Act requires all AI systems generating synthetic text to include machine-readable, detectable watermarks using robust, interoperable technical solutions with two layers. This applies to all AI models, including open-source ones, and extends to any service accessible by EU citizens, regardless of location. Non-compliance risks fines of up to 35 million euros or a percentage of annual income, with providers of 'systemic risk' AI models facing heightened liability.
The paper argues that cultural alignment in NLP requires plural epistemologies, not just diverse data. It proposes a socio-technical model to analyze how multiple, locally grounded ways of knowing can be integrated into language technology, emphasizing that current approaches often fail to address deeper issues of power and governance.
π-RAG decouples LLMs from sensitive data by using π's digits as an immutable, uneditable source of entropy. It introduces a semantic quantization layer that maps user inputs to canonical intent centroids, then uses cryptographic salt to generate deterministic offsets pointing to standardized payloads, ensuring oblivious retrieval and mathematical guarantees of data privacy.
A user reports their Hugging Face account, AntixStudioDesign, was locked unexpectedly during experimentation with AI tools. They have contacted the Safety Team via email and seek advice on account recovery, response time, and data preservation options.
OTTER is a black-box red-teaming framework that bypasses toxicity filters by modifying as few as five tokens. Evaluated on 457 AdvBench prompts across four GPT models, it increases jailbreak success rate from 7.0% to 84.0%, offering the first quantitative analysis of toxicity-bypass relationships and actionable recommendations for classifier hardening.
A validation-gated framework evaluates LLM internal features only after observed behavior, revealing a mid-network feature that causally contributes to suicide detection. This feature is semantic, low-rank, cross-model, and specific to suicidality over general distress, though steering is necessary but not sufficient. The pattern shows smaller models encode suicidality but only larger ones act on it, with evidence limited to English Reddit text.
A new study reveals over 1,000 legal filings contain fabricated citations, with the number rising annually. Benchmarking five AI models shows improved performance, with GPT-5 achieving 82.8% recall and 60.5% F1 in agentic settings, though all models struggle with subtle errors and face resource constraints due to limited information access.
MedLayXPlain introduces the first large-scale benchmark for medical lay language generation, featuring 122,789 region-grounded samples across eight imaging modalities. It evaluates medical vision-language models on expert-lay alignment using a hierarchical ontology system and a lightweight evaluator, revealing a systematic gap: expert-level performance in captioning coexists with significant degradation in lay language, while general-purpose models lack clinical precision.
LISE decomposes speaker embeddings into interpretable components without annotations. Listening experiments show human participants correctly distinguish speakers with 83.9% accuracy, validating the interpretability of the components while preserving ASV performance.
A study finds that sexualised AI voices on commercial platforms reinforce binary gender norms. Female-coded voices are more often described with submissive, sexualised terms, while male-coded voices are linked to dominance and positive traits, reflecting entrenched gendered power asymmetries.
Researchers identify 'role confusion' as a key vulnerability in LLMs, where models misinterpret user input due to stylistic similarities with internal role tags. Destyling user prompts reduces attack success from 61% to 10%, showing that subtle text style changes can dramatically alter model behavior, even when the content appears identical to humans.
Zico Kolter and Matt Fredrikson, co-authors of the definitive paper on indirect prompt injections and authorities on the Mythos model, discuss the growing risks of AI security. They highlight that AI systems require a distinct security mindset, with agents introducing new vulnerabilities, and that specialized red-teaming AI can outperform humans in breaking models, making AI prompt injection breaches increasingly likely.
OpenAI has introduced Codex Security and GPT-5.5-Cyber as part of its Daybreak suite. These tools aim to help organizations identify, validate, and patch vulnerabilities at scale.
NVIDIA has introduced Halos for Robotics, a full-stack functional safety system designed for physical AI. It enables AI-driven safety in unstructured environments where robots operate autonomously alongside humans in factories, warehouses, hospitals, and homes.