WARNING: This paper contains unsafe content. With the development of large language models (LLMs), the sequence length of these models continues to increase, drawing significant attention to long-context language models. However, the evaluation of these models has been primarily limited to their capabilities, with a lack of research focusing on their safety. Existing work, such as ManyShotJailbreak, has to some extent demonstrated that longcontext language models can exhibit safety concerns. However, the methods used are limited and lack comprehensiveness. In response, we introduce LongSafety-Bench, the first benchmark designed to objectively and comprehensively evaluate the safety of long-context models. LongSafetyBench consists of 10 task categories, with an average length of 41,889 words. After testing eight long-context language models on LongSafetyBench, we found that existing models generally exhibit insufficient safety capabilities. The proportion of safe responses from most mainstream long-context LLMs is below 50%. Moreover, models' safety performance in long-context scenarios does not always align with that in short-context scenarios. Further investigation revealed that long-context models tend to overlook harmful content within lengthy texts. We also proposed a simple yet effective solution, allowing open-source models to achieve performance comparable to that of top-tier closed-source models. We believe that LongSafetyBench can serve as a valuable benchmark for evaluating the safety capabilities of long-context language models. We hope that our work will encourage the broader community to pay attention to the safety of long-context models and contribute to the development of solutions to improve the safety of long-context LLMs. Recently, thanks to more advanced model architectures (Xiao et al., 2024b;a; Liu et al., 2024a) and expanded position encoding techniques (Su et al., 2023; Liu et al., 2024b), the context length of language models has been extended significantly (Achiam et al., 2023; Reid et al., 2024). In the foreseeable future, as language models continue to evolve and tackle increasingly complex problems, the demand for handling longer contexts is expected to grow accordingly. We anticipate that long-context language models will become mainstream. Previous research on long-context language models, such as LongBench (Bai et al., 2024), L-Eval (An et al., 2023), and RULER (Hsieh et al., 2024), has typically focused on their capabilities, while neglecting to address their safety. In short-context scenarios, the safety issues of language models have already been extensively studied.(Zhang Illegal Activities, Misinformation Harm, Offensiveness The question is composed of a long content and Bias.

Threat hunting analyzes large, noisy, high-dimensional data to find sparse adversarial behavior. We believe adversarial activities, however they are disguised, are extremely difficult to completely obscure in high dimensional space. In this paper, we employ these latent features of cyber data to find anomalies via a prototype tool called Cyber Log Embeddings Model (CLEM). CLEM was trained on Zeek network traffic logs from both a real-world production network and an from Internet of Things (IoT) cybersecurity testbed. The model is deliberately overtrained on a sliding window of data to characterize each window closely. We use the Adjusted Rand Index (ARI) to comparing the k-means clustering of CLEM output to expert labeling of the embeddings. Our approach demonstrates that there is promise in using natural language modeling to understand cyber data.

Using Privacy-Enhancing Technologies (PETs) for machine learning often influences the characteristics of a machine learning approach, e.g., the needed computational power, timing of the answers or how the data can be utilized. When designing a new service, the developer faces the problem that some decisions require a trade-off. For example, the use of a PET may cause a delay in the responses or adding noise to the data to improve the users' privacy might have a negative impact on the accuracy of the machine learning approach. As of now, there is no structured way how the users' perception of a machine learning based service can contribute to the selection of Privacy Preserving Machine Learning (PPML) methods. This is especially a challenge since one cannot assume that users have a deep technical understanding of these technologies. Therefore, they can only be asked about certain attributes that they can perceive when using the service and not directly which PPML they prefer. This study introduces a decision support framework with the aim of supporting the selection of PPML technologies based on user preferences. Based on prior work analysing User Acceptance Criteria (UAC), we translate these criteria into differentiating characteristics for various PPML techniques. As a final result, we achieve a technology ranking based on the User Acceptance Criteria while providing technology insights for the developers. We demonstrate its application using the use case of classifying privacy-relevant information. Our contribution consists of the decision support framework which consists of a process to connect PPML technologies with UAC, a process for evaluating the characteristics that separate PPML techniques, and a ranking method to evaluate the best PPML technique for the use case.

Artificial intelligence (AI) in medical device software (MDSW) represents a transformative clinical technology, attracting increasing attention within both the medical community and the regulators. In this study, we leverage a data-driven approach to automatically extract and analyze AI-enabled medical devices (AIMD) from the National Medical Products Administration (NMPA) regulatory database. The continued increase in publicly available regulatory data requires scalable methods for analysis. Automation of regulatory information screening is essential to create reproducible insights that can be quickly updated in an ever changing medical device landscape. More than 4 million entries were assessed, identifying 2,174 MDSW registrations, including 531 standalone applications and 1,643 integrated within medical devices, of which 43 were AI-enabled. It was shown that the leading medical specialties utilizing AIMD include respiratory (20.5%), ophthalmology/endocrinology (12.8%), and orthopedics (10.3%). This approach greatly improves the speed of data extracting providing a greater ability to compare and contrast. This study provides the first extensive, data-driven exploration of AIMD in China, showcasing the potential of automated regulatory data analysis in understanding and advancing the landscape of AI in medical technology.

Scaling model sizes to scale performance has worked remarkably well for the current large language models paradigm. The research and empirical findings of various scaling studies led to novel scaling results and laws that guides subsequent research. High training costs for contemporary scales of data and models result in a lack of thorough understanding of how to tune and arrive at such training setups. One direction to ameliorate the cost of pretraining large models is to warmstart the large-scale training from smaller models that are cheaper to tune. In this work, we attempt to understand if the behavior of optimal hyperparameters can be retained under warmstarting for scaling. We explore simple operations that allow the application of theoretically motivated methods of zero-shot transfer of optimal hyperparameters using {\mu}Transfer. We investigate the aspects that contribute to the speedup in convergence and the preservation of stable training dynamics under warmstarting with {\mu}Transfer. We find that shrinking smaller model weights, zero-padding, and perturbing the resulting larger model with scaled initialization from {\mu}P enables effective warmstarting of $\mut{}$.

As large language models (LLMs) grow more powerful, ensuring their safety against misuse becomes crucial. While researchers have focused on developing robust defenses, no method has yet achieved complete invulnerability to attacks. We propose an alternative approach: instead of seeking perfect adversarial robustness, we develop rapid response techniques to look to block whole classes of jailbreaks after observing only a handful of attacks. To study this setting, we develop RapidResponseBench, a benchmark that measures a defense's robustness against various jailbreak strategies after adapting to a few observed examples. We evaluate five rapid response methods, all of which use jailbreak proliferation, where we automatically generate additional jailbreaks similar to the examples observed. Our strongest method, which fine-tunes an input classifier to block proliferated jailbreaks, reduces attack success rate by a factor greater than 240 on an in-distribution set of jailbreaks and a factor greater than 15 on an out-of-distribution set, having observed just one example of each jailbreaking strategy. Moreover, further studies suggest that the quality of proliferation model and number of proliferated examples play an key role in the effectiveness of this defense. Overall, our results highlight the potential of responding rapidly to novel jailbreaks to limit LLM misuse. As Large Language Models (LLMs) become more capable, they pose greater misuse risks. Indeed, the potential for catastrophic misuse of LLMs has motivated AI labs to make public commitments to developing safeguards to minimize the risk of such misuse (Anthropic, 2023; OpenAI, 2023). Additionally, such concerns have motivated substantial effort from the research community to defend against jailbreaks, which are techniques that extract harmful information from LLMs trained to be helpful, harmless, and honest (Bai et al., 2022b; Xie et al., 2023; Xu et al., 2024). Despite ongoing research, ensuring that large language models (LLMs) are robustly resistant to jailbreaking remains an unsolved challenge (Hendrycks et al., 2021; Ziegler et al., 2022). Even state-of-the-art methods that substantially improve robustness, such as representation rerouting (Zou et al., 2024), have been publicly broken within hours of release. The situation could worryingly parallel that of adversarial robustness in computer vision, where new defenses are often defeated by attacks available before their development with proper tuning (Tramer et al., 2020). Indeed, in computer vision, a decade of work and thousands of papers have yielded "limited progress" (Carlini, 2024).

Extended human presence beyond low-Earth orbit (BLEO) during missions to the Moon and Mars will pose significant challenges in the near future. A primary health risk associated with these missions is radiation exposure, primarily from galatic cosmic rays (GCRs) and solar proton events (SPEs). While GCRs present a more consistent, albeit modulated threat, SPEs are harder to predict and can deliver acute doses over short periods. Currently NASA utilizes analytical tools for monitoring the space radiation environment in order to make decisions of immediate action to shelter astronauts. However this reactive approach could be significantly enhanced by predictive models that can forecast radiation exposure in advance, ideally hours ahead of major events, while providing estimates of prediction uncertainty to improve decision-making. In this work we present a machine learning approach for forecasting radiation exposure in BLEO using multimodal time-series data including direct solar imagery from Solar Dynamics Observatory, X-ray flux measurements from GOES missions, and radiation dose measurements from the BioSentinel satellite that was launched as part of Artemis 1 mission. To our knowledge, this is the first time full-disk solar imagery has been used to forecast radiation exposure. We demonstrate that our model can predict the onset of increased radiation due to an SPE event, as well as the radiation decay profile after an event has occurred.

With the widespread of digital environments, reliable authentication and continuous access control has become crucial. It can minimize cyber attacks and prevent frauds, specially those associated with identity theft. A particular interest lies on keystroke dynamics (KD), which refers to the task of recognizing individuals' identity based on their unique typing style. In this work, we propose the use of pre-trained language models (PLMs) to recognize such patterns. Although PLMs have shown high performance on multiple NLP benchmarks, the use of these models on specific tasks requires customization. BERT and RoBERTa, for instance, rely on subword tokenization, and they cannot be directly applied to KD, which requires temporal-character information to recognize users. Recent character-aware PLMs are able to process both subwords and character-level information and can be an alternative solution. Notwithstanding, they are still not suitable to be directly fine-tuned for KD as they are not optimized to account for user's temporal typing information (e.g., hold time and flight time). To overcome this limitation, we propose TempCharBERT, an architecture that incorporates temporal-character information in the embedding layer of CharBERT. This allows modeling keystroke dynamics for the purpose of user identification and authentication. Our results show a significant improvement with this customization. We also showed the feasibility of training TempCharBERT on a federated learning settings in order to foster data privacy.

Artificial Intelligence (AI) has emerged in popularity recently, recording great progress in various industries. However, the environmental impact of AI is a growing concern, in terms of the energy consumption and carbon footprint of Machine Learning (ML) and Deep Learning (DL) models, making essential investigate Green AI, an attempt to reduce the climate impact of AI systems. This paper presents an assessment of different programming languages and Feature Selection (FS) methods to improve computation performance of AI focusing on Network Intrusion Detection (NID) and cyber-attack classification tasks. Experiments were conducted using five ML models - Random Forest, XGBoost, LightGBM, Multi-Layer Perceptron, and Long Short-Term Memory - implemented in four programming languages - Python, Java, R, and Rust - along with three FS methods - Information Gain, Recursive Feature Elimination, and Chi-Square. The obtained results demonstrated that FS plays an important role enhancing the computational efficiency of AI models without compromising detection accuracy, highlighting languages like Python and R, that benefit from a rich AI libraries environment. These conclusions can be useful to design efficient and sustainable AI systems that still provide a good generalization and a reliable detection.

Instruction tuning has been widely adopted to ensure large language models (LLMs) follow user instructions effectively. The resulting instruction-following capabilities of LLMs heavily rely on the instruction datasets used for tuning. Recently, synthetic instruction datasets have emerged as an economically viable solution to provide LLMs diverse and high-quality instructions. However, existing approaches typically assume that larger or stronger models are stronger teachers for instruction tuning, and hence simply adopt these models as response generators to the synthetic instructions. In this paper, we challenge this commonly-adopted assumption. Our extensive experiments across five base models and twenty response generators reveal that larger and stronger models are not necessarily stronger teachers of smaller models. We refer to this phenomenon as the Larger Models' Paradox. We observe that existing metrics cannot precisely predict the effectiveness of response generators since they ignore the compatibility between teachers and base models being fine-tuned. We thus develop a novel metric, named as Compatibility-Adjusted Reward (CAR) to measure the effectiveness of response generators. Our experiments across five base models demonstrate that CAR outperforms almost all baselines.