Black-box Vulnerabilities

The Automated Progressive Red Teaming (APRT) framework exploits vulnerabilities in large language models (LLMs) by iteratively generating adversarial prompts. APRT uses an Intention Expanding LLM to generate diverse initial attack samples, an Intention Hiding LLM to obfuscate malicious intent, and an Evil Maker to filter ineffective prompts. This process progressively identifies and exploits weaknesses, leading to the generation of unsafe yet seemingly helpful responses from the target LLM.

The SoP framework allows for automated generation of jailbreak prompts, bypassing safety mechanisms in LLMs. SoP utilizes multiple automatically optimized "jailbreak characters" within a single prompt to persuade the LLM to generate harmful or undesirable content, even without any seed jailbreak templates. This vulnerability is demonstrated against GPT-3.5, GPT-4, and LLaMA-2.

Appending a single whitespace character (space) or certain punctuation marks to the end of an LLM's input template can bypass safety mechanisms and cause the model to generate unsafe, biased, or factually incorrect outputs, even if the original prompt was benign. This vulnerability is due to the statistical properties of single-character tokens in the model's training data, causing unintended behavior in the model's token prediction.

Large language models (LLMs) employing safety measures like filters and alignment training remain vulnerable to information leakage via "Decomposition Attacks". These attacks decompose a malicious query into multiple benign sub-queries, eliciting responses from the LLM that, when aggregated, reveal sensitive information without triggering safety filters or producing directly harmful outputs.

Large Language Models (LLMs) exhibit vulnerability to jailbreak attacks when prompted using Arabic transliteration and Arabizi (Arabic chatspeak). While LLMs demonstrate robustness to standard Arabic prompts, even with prefix injection, the use of transliterated or Arabizi prompts bypasses safety mechanisms, leading to the generation of unsafe content. This is due to the model's learned associations with specific words in these non-standard forms, which differ from its understanding of the…

Large Vision Language Models (LVLMs) are vulnerable to a bi-modal adversarial prompt attack (BAP). BAP leverages a combined textual and visual prompt to bypass safety mechanisms and elicit harmful responses, even in models designed to resist single-modality attacks. The attack first introduces a query-agnostic adversarial perturbation to the visual prompt, making the model more likely to respond positively regardless of the text. Then, an LLM refines the textual prompt iteratively to achieve…

Large Language Models (LLMs) are vulnerable to a black-box query-response optimization attack (QROA). QROA iteratively refines a malicious prompt suffix using a surrogate model to maximize a reward function that measures the likelihood of eliciting harmful content from the LLM. This attack does not require access to the model's internal parameters or logits; it operates solely via standard query-response interactions.

Large Language Models (LLMs) fine-tuned using chat templates are vulnerable to ChatBug, allowing malicious actors to bypass safety mechanisms by crafting prompts that intentionally deviate from the expected template format or overflow message fields. This exploits the LLM’s reliance on the template structure without enforcing similar constraints on user input.

Large Language Models (LLMs) exhibit increased vulnerability to adversarial prompts employing code-switching techniques, where multiple languages are interwoven within a single query. This vulnerability stems from an unintended correlation between the resource availability of the languages used in the prompt and the LLM's safety alignment. LLMs trained on imbalanced multilingual data are more susceptible to attacks leveraging low-resource languages, resulting in a higher rate of unsafe or…

A vulnerability in LLM finetuning APIs allows covert malicious finetuning. Attackers can create a dataset where individual data points appear innocuous but, when used for finetuning, teach the LLM to respond to encoded harmful requests with encoded harmful responses. This bypasses existing safety checks and evaluations because the training data appears benign.