Public LLMs such as the Llama 2-Chat underwent alignment training and were considered safe. Recently Qi et al. [2024] reported that even benign fine-tuning on seemingly safe datasets can give rise to unsafe behaviors in the models.

The robustness and security of large language models (LLMs) has become a prominent research area. One notable vulnerability is the ability to bypass LLM safeguards by translating harmful queries into rare or underrepresented languages, a simple yet effective method of "jailbreaking" these models. Despite the growing concern, there has been limited research addressing the safeguarding of LLMs in multilingual scenarios, highlighting an urgent need to enhance multilingual safety. In this work, we investigate the correlation between various attack features across different languages and propose Multilingual Collaborative Defense (MCD), a novel learning method that optimizes a continuous, soft safety prompt automatically to facilitate multilingual safeguarding of LLMs. The MCD approach offers three advantages: First, it effectively improves safeguarding performance across multiple languages. Second, MCD maintains strong generalization capabilities while minimizing false refusal rates. Third, MCD mitigates the language safety misalignment caused by imbalances in LLM training corpora. To evaluate the effectiveness of MCD, we manually construct multilingual versions of commonly used jailbreak benchmarks, such as MaliciousInstruct and AdvBench, to assess various safeguarding methods. Additionally, we introduce these datasets in underrepresented (zero-shot) languages to verify the language transferability of MCD. The results demonstrate that MCD outperforms existing approaches in safeguarding against multilingual jailbreak attempts while also exhibiting strong language transfer capabilities. Our code is available at https://github.com/HLiang-Lee/MCD.

Large Vision-Language Models (LVLMs) exhibit impressive potential across various tasks but also face significant privacy risks, limiting their practical applications. Current researches on privacy assessment for LVLMs is limited in scope, with gaps in both assessment dimensions and privacy categories. To bridge this gap, we propose Multi-P$^2$A, a comprehensive benchmark for evaluating the privacy preservation capabilities of LVLMs in terms of privacy awareness and leakage. Privacy awareness measures the model's ability to recognize the privacy sensitivity of input data, while privacy leakage assesses the risk of the model unintentionally disclosing privacy information in its output. We design a range of sub-tasks to thoroughly evaluate the model's privacy protection offered by LVLMs. Multi-P$^2$A covers 26 categories of personal privacy, 15 categories of trade secrets, and 18 categories of state secrets, totaling 31,962 samples. Based on Multi-P$^2$A, we evaluate the privacy preservation capabilities of 21 open-source and 2 closed-source LVLMs. Our results reveal that current LVLMs generally pose a high risk of facilitating privacy breaches, with vulnerabilities varying across personal privacy, trade secret, and state secret.

Large Language Models (LLMs) have demonstrated exceptional capabilities across various natural language processing tasks. Due to their training on internet-sourced datasets, LLMs can sometimes generate objectionable content, necessitating extensive alignment with human feedback to avoid such outputs. Despite massive alignment efforts, LLMs remain susceptible to adversarial jailbreak attacks, which usually are manipulated prompts designed to circumvent safety mechanisms and elicit harmful responses. Here, we introduce a novel approach, Directed Rrepresentation Optimization Jailbreak (DROJ), which optimizes jailbreak prompts at the embedding level to shift the hidden representations of harmful queries towards directions that are more likely to elicit affirmative responses from the model. Our evaluations on LLaMA-2-7b-chat model show that DROJ achieves a 100\% keyword-based Attack Success Rate (ASR), effectively preventing direct refusals. However, the model occasionally produces repetitive and non-informative responses. To mitigate this, we introduce a helpfulness system prompt that enhances the utility of the model's responses. Our code is available at https://github.com/Leon-Leyang/LLM-Safeguard.

Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious instruction prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful responses from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness. This paper examines the LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Our novel decoder-oriented, step-bystep defense architecture corrects harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost secure decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods. In recent years, significant progress has been made in developing large language models (LLMs). Meanwhile, the safety of LLMs has attracted significant attention from the research community and industry (Weidinger et al., 2021; Achiam et al., 2023; Wu et al., 2023b). One of the primary safety concerns is jailbreaking, where malicious actors or errant inputs prompt LLMs to produce harmful or inappropriate content, effectively bypassing ethical guidelines.

Public LLMs such as the Llama 2-Chat have driven huge activity in LLM research. These models underwent alignment training and were considered safe. Recently Qi et al. (2023) reported that even benign fine-tuning (e.g., on seemingly safe datasets) can give rise to unsafe behaviors in the models. The current paper is about methods and best practices to mitigate such loss of alignment. Through extensive experiments on several chat models (Meta's Llama 2-Chat, Mistral AI's Mistral 7B Instruct v0.2, and OpenAI's GPT-3.5 Turbo), this paper uncovers that the prompt templates used during fine-tuning and inference play a crucial role in preserving safety alignment, and proposes the "Pure Tuning, Safe Testing" (PTST) principle -- fine-tune models without a safety prompt, but include it at test time. Fine-tuning experiments on GSM8K, ChatDoctor, and OpenOrca show that PTST significantly reduces the rise of unsafe behaviors, and even almost eliminates them in some cases.

Prepending model inputs with safety prompts is a common practice of safeguarding large language models (LLMs) from complying with queries that contain harmful intents. However, the working mechanisms of safety prompts have not yet been fully understood, which hinders the potential for automatically optimizing them for improved LLM safety. Motivated by this problem, we investigate the impact of safety prompts from the perspective of model representations. We find that in models' representation space, harmful and harmless queries can be largely distinguished, but this is not noticeably enhanced by safety prompts. Instead, the queries' representations are moved by different safety prompts in similar directions, where models become more prone to refusal (i.e., refusing to provide assistance) even when the queries are harmless. Inspired by these findings, we propose a method called DRO (Directed Representation Optimization) for automatic safety prompt optimization. DRO treats safety prompts as continuous, trainable embeddings and learns to move the representations of harmful/harmless queries along/opposite the direction in which the model's refusal probability increases. We demonstrate that DRO remarkably improves the safeguarding performance of human-crafted safety prompts and outperforms strong baselines, as evaluated on out-of-domain benchmarks, without compromising the general model capability.

With the rapid evolution of large language models (LLMs), there is a growing concern that they may pose risks or have negative social impacts. Therefore, evaluation of human values alignment is becoming increasingly important. Previous work mainly focuses on assessing the performance of LLMs on certain knowledge and reasoning abilities, while neglecting the alignment to human values, especially in a Chinese context. In this paper, we present CValues, the first Chinese human values evaluation benchmark to measure the alignment ability of LLMs in terms of both safety and responsibility criteria. As a result, we have manually collected adversarial safety prompts across 10 scenarios and induced responsibility prompts from 8 domains by professional experts. To provide a comprehensive values evaluation of Chinese LLMs, we not only conduct human evaluation for reliable comparison, but also construct multi-choice prompts for automatic evaluation. Our findings suggest that while most Chinese LLMs perform well in terms of safety, there is considerable room for improvement in terms of responsibility. Moreover, both the automatic and human evaluation are important for assessing the human values alignment in different aspects. The benchmark and code is available on ModelScope and Github.