Research Papers

StarCoder: may the source be with you!

Raymond Li, Loubna Ben Allal, Yangtian Zi +64

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.

Most traditional AI safety research has approached AI models as machines and centered on algorithm-focused attacks developed by security experts. As large language models (LLMs) become increasingly common and competent, non-expert users can also impose risks during daily interactions. This paper introduces a new perspective to jailbreak LLMs as human-like communicators, to explore this overlooked intersection between everyday language interaction and AI safety. Specifically, we study how to persuade LLMs to jailbreak them. First, we propose a persuasion taxonomy derived from decades of social science research. Then, we apply the taxonomy to automatically generate interpretable persuasive adversarial prompts (PAP) to jailbreak LLMs. Results show that persuasion significantly increases the jailbreak performance across all risk categories: PAP consistently achieves an attack success rate of over $92\%$ on Llama 2-7b Chat, GPT-3.5, and GPT-4 in $10$ trials, surpassing recent algorithm-focused attacks. On the defense side, we explore various mechanisms against PAP and, found a significant gap in existing defenses, and advocate for more fundamental mitigation for highly interactive LLMs

The rapid development of Artificial Intelligence (AI) technology has enabled the deployment of various systems based on it. However, many current AI systems are found vulnerable to imperceptible attacks, biased against underrepresented groups, lacking in user privacy protection. These shortcomings degrade user experience and erode people's trust in all AI systems. In this review, we provide AI practitioners with a comprehensive guide for building trustworthy AI systems. We first introduce the theoretical framework of important aspects of AI trustworthiness, including robustness, generalization, explainability, transparency, reproducibility, fairness, privacy preservation, and accountability. To unify currently available but fragmented approaches toward trustworthy AI, we organize them in a systematic approach that considers the entire lifecycle of AI systems, ranging from data acquisition to model development, to system development and deployment, finally to continuous monitoring and governance. In this framework, we offer concrete action items for practitioners and societal stakeholders (e.g., researchers, engineers, and regulators) to improve AI trustworthiness. Finally, we identify key opportunities and challenges for the future development of trustworthy AI systems, where we identify the need for a paradigm shift toward comprehensively trustworthy AI systems.

The robustness of LLMs to jailbreak attacks, where users design prompts to circumvent safety measures and misuse model capabilities, has been studied primarily for LLMs acting as simple chatbots. Meanwhile, LLM agents -- which use external tools and can execute multi-stage tasks -- may pose a greater risk if misused, but their robustness remains underexplored. To facilitate research on LLM agent misuse, we propose a new benchmark called AgentHarm. The benchmark includes a diverse set of 110 explicitly malicious agent tasks (440 with augmentations), covering 11 harm categories including fraud, cybercrime, and harassment. In addition to measuring whether models refuse harmful agentic requests, scoring well on AgentHarm requires jailbroken agents to maintain their capabilities following an attack to complete a multi-step task. We evaluate a range of leading LLMs, and find (1) leading LLMs are surprisingly compliant with malicious agent requests without jailbreaking, (2) simple universal jailbreak templates can be adapted to effectively jailbreak agents, and (3) these jailbreaks enable coherent and malicious multi-step agent behavior and retain model capabilities. To enable simple and reliable evaluation of attacks and defenses for LLM-based agents, we publicly release AgentHarm at https://huggingface.co/datasets/ai-safety-institute/AgentHarm.

Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack

Mark Russinovich, Ahmed Salem, Ronen Eldan

Large Language Models (LLMs) have risen significantly in popularity and are increasingly being adopted across multiple applications. These LLMs are heavily aligned to resist engaging in illegal or unethical topics as a means to avoid contributing to responsible AI harms. However, a recent line of attacks, known as jailbreaks, seek to overcome this alignment. Intuitively, jailbreak attacks aim to narrow the gap between what the model can do and what it is willing to do. In this paper, we introduce a novel jailbreak attack called Crescendo. Unlike existing jailbreak methods, Crescendo is a simple multi-turn jailbreak that interacts with the model in a seemingly benign manner. It begins with a general prompt or question about the task at hand and then gradually escalates the dialogue by referencing the model's replies progressively leading to a successful jailbreak. We evaluate Crescendo on various public systems, including ChatGPT, Gemini Pro, Gemini-Ultra, LlaMA-2 70b and LlaMA-3 70b Chat, and Anthropic Chat. Our results demonstrate the strong efficacy of Crescendo, with it achieving high attack success rates across all evaluated models and tasks. Furthermore, we present Crescendomation, a tool that automates the Crescendo attack and demonstrate its efficacy against state-of-the-art models through our evaluations. Crescendomation surpasses other state-of-the-art jailbreaking techniques on the AdvBench subset dataset, achieving 29-61% higher performance on GPT-4 and 49-71% on Gemini-Pro. Finally, we also demonstrate Crescendo's ability to jailbreak multimodal models.

Low-Resource Languages Jailbreak GPT-4

Zheng-Xin Yong, Cristina Menghini, Stephen H. Bach

AI safety training and red-teaming of large language models (LLMs) are measures to mitigate the generation of unsafe content. Our work exposes the inherent cross-lingual vulnerability of these safety mechanisms, resulting from the linguistic inequality of safety training data, by successfully circumventing GPT-4's safeguard through translating unsafe English inputs into low-resource languages. On the AdvBenchmark, GPT-4 engages with the unsafe translated inputs and provides actionable items that can get the users towards their harmful goals 79% of the time, which is on par with or even surpassing state-of-the-art jailbreaking attacks. Other high-/mid-resource languages have significantly lower attack success rate, which suggests that the cross-lingual vulnerability mainly applies to low-resource languages. Previously, limited training on low-resource languages primarily affects speakers of those languages, causing technological disparities. However, our work highlights a crucial shift: this deficiency now poses a risk to all LLMs users. Publicly available translation APIs enable anyone to exploit LLMs' safety vulnerabilities. Therefore, our work calls for a more holistic red-teaming efforts to develop robust multilingual safeguards with wide language coverage.

Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models

Xianjun Yang, Xiao Wang, Qi Zhang +4

Warning: This paper contains examples of harmful language, and reader discretion is recommended. The increasing open release of powerful large language models (LLMs) has facilitated the development of downstream applications by reducing the essential cost of data annotation and computation. To ensure AI safety, extensive safety-alignment measures have been conducted to armor these models against malicious use (primarily hard prompt attack). However, beneath the seemingly resilient facade of the armor, there might lurk a shadow. By simply tuning on 100 malicious examples with 1 GPU hour, these safely aligned LLMs can be easily subverted to generate harmful content. Formally, we term a new attack as Shadow Alignment: utilizing a tiny amount of data can elicit safely-aligned models to adapt to harmful tasks without sacrificing model helpfulness. Remarkably, the subverted models retain their capability to respond appropriately to regular inquiries. Experiments across 8 models released by 5 different organizations (LLaMa-2, Falcon, InternLM, BaiChuan2, Vicuna) demonstrate the effectiveness of shadow alignment attack. Besides, the single-turn English-only attack successfully transfers to multi-turn dialogue and other languages. This study serves as a clarion call for a collective effort to overhaul and fortify the safety of open-source LLMs against malicious attackers.

Saving Face: Investigating the Ethical Concerns of Facial Recognition Auditing

Inioluwa Deborah Raji, Timnit Gebru, Margaret Mitchell +3

Although essential to revealing biased performance, well intentioned attempts at algorithmic auditing can have effects that may harm the very populations these measures are meant to protect. This concern is even more salient while auditing biometric systems such as facial recognition, where the data is sensitive and the technology is often used in ethically questionable manners. We demonstrate a set of five ethical concerns in the particular case of auditing commercial facial processing technology, highlighting additional design considerations and ethical tensions the auditor needs to be aware of so as not exacerbate or complement the harms propagated by the audited system. We go further to provide tangible illustrations of these concerns, and conclude by reflecting on what these concerns mean for the role of the algorithmic audit and the fundamental product limitations they reveal.

On the Robustness of ChatGPT: An Adversarial and Out-of-distribution Perspective

Jindong Wang, Xixu Hu, Wenxin Hou +10

ChatGPT is a recent chatbot service released by OpenAI and is receiving increasing attention over the past few months. While evaluations of various aspects of ChatGPT have been done, its robustness, i.e., the performance to unexpected inputs, is still unclear to the public. Robustness is of particular concern in responsible AI, especially for safety-critical applications. In this paper, we conduct a thorough evaluation of the robustness of ChatGPT from the adversarial and out-of-distribution (OOD) perspective. To do so, we employ the AdvGLUE and ANLI benchmarks to assess adversarial robustness and the Flipkart review and DDXPlus medical diagnosis datasets for OOD evaluation. We select several popular foundation models as baselines. Results show that ChatGPT shows consistent advantages on most adversarial and OOD classification and translation tasks. However, the absolute performance is far from perfection, which suggests that adversarial and OOD robustness remains a significant threat to foundation models. Moreover, ChatGPT shows astounding performance in understanding dialogue-related texts and we find that it tends to provide informal suggestions for medical tasks instead of definitive answers. Finally, we present in-depth discussions of possible research directions.

Current machine-learning based software vulnerability detection methods are primarily conducted at the function-level. However, a key limitation of these methods is that they do not indicate the specific lines of code contributing to vulnerabilities. This limits the ability of developers to efficiently inspect and interpret the predictions from a learnt model, which is crucial for integrating machine-learning based tools into the software development workflow. Graph-based models have shown promising performance in function-level vulnerability detection, but their capability for statement-level vulnerability detection has not been extensively explored. While interpreting function-level predictions through explainable AI is one promising direction, we herein consider the statement-level software vulnerability detection task from a fully supervised learning perspective. We propose a novel deep learning framework, LineVD, which formulates statement-level vulnerability detection as a node classification task. LineVD leverages control and data dependencies between statements using graph neural networks, and a transformer-based model to encode the raw source code tokens. In particular, by addressing the conflicting outputs between function-level and statement-level information, LineVD significantly improve the prediction performance without vulnerability status for function code. We have conducted extensive experiments against a large-scale collection of real-world C/C++ vulnerabilities obtained from multiple real-world projects, and demonstrate an increase of 105\% in F1-score over the current state-of-the-art.

Recent technological advances in synthetic data have enabled the generation of images with such high quality that human beings cannot tell the difference between real-life photographs and Artificial Intelligence (AI) generated images. Given the critical necessity of data reliability and authentication, this article proposes to enhance our ability to recognise AI-generated images through computer vision. Initially, a synthetic dataset is generated that mirrors the ten classes of the already available CIFAR-10 dataset with latent diffusion which provides a contrasting set of images for comparison to real photographs. The model is capable of generating complex visual attributes, such as photorealistic reflections in water. The two sets of data present as a binary classification problem with regard to whether the photograph is real or generated by AI. This study then proposes the use of a Convolutional Neural Network (CNN) to classify the images into two categories; Real or Fake. Following hyperparameter tuning and the training of 36 individual network topologies, the optimal approach could correctly classify the images with 92.98% accuracy. Finally, this study implements explainable AI via Gradient Class Activation Mapping to explore which features within the images are useful for classification. Interpretation reveals interesting concepts within the image, in particular, noting that the actual entity itself does not hold useful information for classification; instead, the model focuses on small visual imperfections in the background of the images. The complete dataset engineered for this study, referred to as the CIFAKE dataset, is made publicly available to the research community for future work.

Trustworthy AI

Richa Singh, Mayank Vatsa, Nalini Ratha

Modern AI systems are reaping the advantage of novel learning methods. With their increasing usage, we are realizing the limitations and shortfalls of these systems. Brittleness to minor adversarial changes in the input data, ability to explain the decisions, address the bias in their training data, high opacity in terms of revealing the lineage of the system, how they were trained and tested, and under which parameters and conditions they can reliably guarantee a certain level of performance, are some of the most prominent limitations. Ensuring the privacy and security of the data, assigning appropriate credits to data sources, and delivering decent outputs are also required features of an AI system. We propose the tutorial on Trustworthy AI to address six critical issues in enhancing user and public trust in AI systems, namely: (i) bias and fairness, (ii) explainability, (iii) robust mitigation of adversarial attacks, (iv) improved privacy and security in model building, (v) being decent, and (vi) model attribution, including the right level of credit assignment to the data sources, model architectures, and transparency in lineage.