Existing work investigates the reasoning capabilities of large language models (LLMs) to uncover their limitations, human-like biases and underlying processes. Such studies include evaluations of base LLMs (pre-trained on unlabeled corpora only) for this purpose. Our position paper argues that evaluating base LLMs' reasoning capabilities raises inherent methodological concerns that are overlooked in such existing studies. We highlight the fundamental mismatch between base LLMs' pretraining objective and normative qualities, such as correctness, by which reasoning is assessed. In particular, we show how base LLMs generate logically valid or invalid conclusions as coincidental byproducts of conforming to purely linguistic patterns of statistical plausibility. This fundamental mismatch challenges the assumptions that (a) base LLMs' outputs can be assessed as their bona fide attempts at correct answers or conclusions; and (b) conclusions about base LLMs' reasoning can generalize to post-trained LLMs optimized for successful instruction-following. We call for a critical re-examination of existing work that relies implicitly on these assumptions, and for future work to account for these methodological pitfalls.

These attacks allow an adversary to cause an otherwise "aligned" model--that typically refuses requests such as "how do I build a bomb?" or "swear at me!"--to comply with such requests, or

Log files provide valuable information for detecting and diagnosing problems in enterprise software applications and data centers. Several log analytics tools and platforms were developed to help filter and extract information from logs, typically using regular expressions (RegExes). Recent commercial log analytics platforms provide domain-specific languages specifically designed for log parsing, such as Grok or the Dynatrace Pattern Language (DPL). However, users who want to migrate to these platforms must manually convert their RegExes into the new pattern language, which is costly and error-prone. In this work, we present Reptile, which combines a rule-based approach for converting RegExes into DPL patterns with a best-effort approach for cases where a full conversion is impossible. Furthermore, it integrates GPT-4 to optimize the obtained DPL patterns. The evaluation with 946 RegExes collected from a large company shows that Reptile safely converted 73.7% of them. The evaluation of Reptile's pattern optimization with 23 real-world RegExes showed an F1-score and MCC above 0.91. These results are promising and have ample practical implications for companies that migrate to a modern log analytics platform, such as Dynatrace.

Recent work has shown it is possible to construct adversarial examples that cause aligned language models to emit harmful strings or perform harmful behavior. Existing attacks work either in the white-box setting (with full access to the model weights), or through transferability: the phenomenon that adversarial examples crafted on one model often remain effective on other models. We improve on prior work with a query-based attack that leverages API access to a remote language model to construct adversarial examples that cause the model to emit harmful strings with (much) higher probability than with transfer-only attacks. We validate our attack on GPT-3.5 and OpenAI's safety classifier; we can cause GPT-3.5 to emit harmful strings that current transfer attacks fail at, and we can evade the OpenAI and Llama Guard safety classifiers with nearly 100% probability.

Large Language Model (LLM) services and models often come with legal rules on who can use them and how they must use them. Assessing the compliance of the released LLMs is crucial, as these rules protect the interests of the LLM contributor and prevent misuse. In this context, we describe the novel fingerprinting problem of Black-box Identity Verification (BBIV). The goal is to determine whether a third-party application uses a certain LLM through its chat function. We propose a method called Targeted Random Adversarial Prompt (TRAP) that identifies the specific LLM in use. We repurpose adversarial suffixes, originally proposed for jailbreaking, to get a pre-defined answer from the target LLM, while other models give random answers. TRAP detects the target LLMs with over 95% true positive rate at under 0.2% false positive rate even after a single interaction. TRAP remains effective even if the LLM has minor changes that do not significantly alter the original function.

Large language models (LLMs) exhibit excellent ability to understand human languages, but do they also understand their own language that appears gibberish to us? In this work we delve into this question, aiming to uncover the mechanisms underlying such behavior in LLMs. We employ the Greedy Coordinate Gradient optimizer to craft prompts that compel LLMs to generate coherent responses from seemingly nonsensical inputs. We call these inputs LM Babel and this work systematically studies the behavior of LLMs manipulated by these prompts. We find that the manipulation efficiency depends on the target text's length and perplexity, with the Babel prompts often located in lower loss minima compared to natural prompts. We further examine the structure of the Babel prompts and evaluate their robustness. Notably, we find that guiding the model to generate harmful texts is not more difficult than into generating benign texts, suggesting lack of alignment for out-of-distribution prompts.

Large Language Models (LLMs) have demonstrated remarkable capabilities in various domains, but their vulnerability to trojan or backdoor attacks poses significant security risks. This paper explores the challenges and insights gained from the Trojan Detection Competition 2023 (TDC2023), which focused on identifying and evaluating trojan attacks on LLMs. We investigate the difficulty of distinguishing between intended and unintended triggers, as well as the feasibility of reverse engineering trojans in real-world scenarios. Our comparative analysis of various trojan detection methods reveals that achieving high Recall scores is significantly more challenging than obtaining high Reverse-Engineering Attack Success Rate (REASR) scores. The top-performing methods in the competition achieved Recall scores around 0.16, comparable to a simple baseline of randomly sampling sentences from a distribution similar to the given training prefixes. This finding raises questions about the detectability and recoverability of trojans inserted into the model, given only the harmful targets. Despite the inability to fully solve the problem, the competition has led to interesting observations about the viability of trojan detection and improved techniques for optimizing LLM input prompts. The phenomenon of unintended triggers and the difficulty in distinguishing them from intended triggers highlights the need for further research into the robustness and interpretability of LLMs. The TDC2023 has provided valuable insights into the challenges and opportunities associated with trojan detection in LLMs, laying the groundwork for future research in this area to ensure their safety and reliability in real-world applications.

Large Language Models (LLMs) have surged in popularity in recent months, but they have demonstrated concerning capabilities to generate harmful content when manipulated. While techniques like safety fine-tuning aim to minimize harmful use, recent works have shown that LLMs remain vulnerable to attacks that elicit toxic responses. In this work, we introduce the Proxy-Guided Attack on LLMs (PAL), the first optimization-based attack on LLMs in a black-box query-only setting. In particular, it relies on a surrogate model to guide the optimization and a sophisticated loss designed for real-world LLM APIs. Our attack achieves 84% attack success rate (ASR) on GPT-3.5-Turbo and 48% on Llama-2-7B, compared to 4% for the current state of the art. We also propose GCG++, an improvement to the GCG attack that reaches 94% ASR on white-box Llama-2-7B, and the Random-Search Attack on LLMs (RAL), a strong but simple baseline for query-based attacks. We believe the techniques proposed in this work will enable more comprehensive safety testing of LLMs and, in the long term, the development of better security guardrails. The code can be found at https://github.com/chawins/pal.

Dense video captioning aims to identify the events of interest in an input video, and generate descriptive captions for each event. Previous approaches usually follow a two-stage generative process, which first proposes a segment for each event, then renders a caption for each identified segment. Recent advances in large-scale sequence generation pretraining have seen great success in unifying task formulation for a great variety of tasks, but so far, more complex tasks such as dense video captioning are not able to fully utilize this powerful paradigm. In this work, we show how to model the two subtasks of dense video captioning jointly as one sequence generation task, and simultaneously predict the events and the corresponding descriptions. Experiments on YouCook2 and ViTT show encouraging results and indicate the feasibility of training complex tasks such as end-to-end dense video captioning integrated into large-scale pretrained models.