We study how to subvert large language models (LLMs) from following prompt-specified rules. We first formalize rule-following as inference in propositional Horn logic, a mathematical system in which rules have the form "if $P$ and $Q$, then $R$" for some propositions $P$, $Q$, and $R$. Next, we prove that although small transformers can faithfully follow such rules, maliciously crafted prompts can still mislead both theoretical constructions and models learned from data. Furthermore, we demonstrate that popular attack algorithms on LLMs find adversarial prompts and induce attention patterns that align with our theory. Our novel logic-based framework provides a foundation for studying LLMs in rule-based settings, enabling a formal analysis of tasks like logical reasoning and jailbreak attacks.

Security vulnerabilities in modern software are prevalent and harmful. While automated vulnerability detection techniques have made promising progress, their scalability and applicability remain challenging. The remarkable performance of Large Language Models (LLMs), such as GPT-4 and CodeLlama, on code-related tasks has prompted recent works to explore if LLMs can be used to detect security vulnerabilities. In this paper, we perform a more comprehensive study by examining a larger and more diverse set of datasets, languages, and LLMs, and qualitatively evaluating detection performance across prompts and vulnerability classes. Concretely, we evaluate the effectiveness of 16 pre-trained LLMs on 5,000 code samples-1,000 randomly selected each from five diverse security datasets. These balanced datasets encompass synthetic and real-world projects in Java and C/C++ and cover 25 distinct vulnerability classes. Our results show that LLMs across all scales and families show modest effectiveness in end-to-end reasoning about vul-nerabilities, obtaining an average accuracy of 62.8% and F1 score of 0.71 across all datasets. LLMs are significantly better at detecting vulnerabilities that typically only need intra-procedural reasoning, such as OS Command Injection and NULL Pointer Dereference. Moreover, LLMs report higher accuracies on these vulnerabilities than popular static analysis tools, such as CodeQL. We find that advanced prompting strategies that involve step-by-step analysis significantly improve performance of LLMs on real-world datasets in terms of F1 score (by up to 0.18 on average). Interestingly, we observe that LLMs show promising abilities at performing parts of the analysis correctly, such as identifying vulnerability-related specifications (e.g., sources and sinks) and leveraging natural language information to understand code behavior (e.g., to check if code is sanitized). We believe our insights can motivate future work on LLM-augmented vulnerability detection systems.

Developers spend a significant amount of time in editing code for a variety of reasons such as bug fixing or adding new features. Designing effective methods to predict code edits has been an active yet challenging area of research due to the diversity of code edits and the difficulty of capturing the developer intent. In this work, we address these challenges by endowing pre-trained large language models (LLMs) with the knowledge of relevant prior associated edits, which we call the Grace (Generation conditioned on Associated Code Edits) method. The generative capability of the LLMs helps address the diversity in code changes and conditioning code generation on prior edits helps capture the latent developer intent. We evaluate two well-known LLMs, codex and CodeT5, in zero-shot and fine-tuning settings respectively. In our experiments with two datasets, Grace boosts the performance of the LLMs significantly, enabling them to generate 29% and 54% more correctly edited code in top-1 suggestions relative to the current state-of-the-art symbolic and neural approaches, respectively.