Integrity Vulnerabilities

Large Reasoning Models (LRMs) are vulnerable to a novel jailbreak attack, "Mousetrap," which leverages the models' reasoning capabilities to elicit harmful responses. Mousetrap uses a "Chaos Machine" to iteratively transform prompts via one-to-one mappings (e.g., character substitutions, word reversals), creating complex reasoning chains that confuse the LRM and cause it to generate unsafe outputs despite safety mechanisms. The iterative nature of the attack, combined with role-playing…

A vulnerability exists in Large Language Models (LLMs) that allows for efficient jailbreaking by selectively fine-tuning only the lower layers of the model with a toxic dataset. This "Freeze Training" method, as described in the research paper, concentrates the fine-tuning on layers identified as being highly sensitive to the generation of harmful content. This approach significantly reduces training duration and GPU memory consumption while maintaining a high jailbreak success rate.

Large Language Models (LLMs) with refusal training are vulnerable to a "jailbreaking-to-jailbreak" (J2) attack. A J2 attack involves initially jailbreaking a powerful LLM to create a "J2 attacker." This attacker, instructed with general jailbreaking strategies, then autonomously attempts to jailbreak other LLMs, including potentially the same model it was derived from, by iteratively refining its attack based on previous attempts and in-context learning.

Large Language Models (LLMs) are vulnerable to structure transformation attacks, where malicious prompts are encoded in diverse syntax spaces (e.g., SQL, JSON, LLM-generated syntaxes) to bypass safety mechanisms. These attacks maintain the harmful intent while altering the linguistic structure, making detection based on token-level patterns ineffective.

Large Language Model (LLM) watermarking schemes based on n-gram probability biases (specifically KGW, SynthID-Text, MinHash, and SkipHash) are vulnerable to adversarial removal during Knowledge Distillation. When a student model is trained on the output of a watermarked teacher model, it inherits the watermark's statistical biases ("radioactivity"). An attacker can exploit this inheritance by comparing the student model's output token probabilities against a base model to extract the…

Large Language Models (LLMs) are vulnerable to "Rewrite to Jailbreak" (R2J) attacks. R2J exploits the models' safety mechanisms by iteratively rewriting harmful prompts, subtly altering wording to bypass safety filters while maintaining the original malicious intent. This differs from previous methods which rely on adding extraneous prefixes/suffixes or creating forced instruction-following scenarios, thus being more difficult to detect.

Large Language Models (LLMs) trained with safety fine-tuning techniques are vulnerable to multi-dimensional evasion attacks. Safety-aligned behavior, such as refusing harmful queries, is controlled not by a single direction in activation space, but by a subspace of interacting directions. Manipulating non-dominant directions, which represent distinct jailbreak patterns or indirect features, can suppress the dominant direction responsible for refusal, thereby bypassing learned safety…

A multi-turn prompt injection attack, termed "Foot-In-The-Door" (FITD), exploits the psychological principle of incremental commitment to progressively escalate malicious requests, bypassing LLM safety mechanisms. The attack leverages intermediate "bridge" prompts and self-alignment techniques to coax the model into generating increasingly harmful outputs, even when initially refusing similar direct requests.

A novel "Flanking Attack" exploits the vulnerability of multimodal LLMs (e.g., Google Gemini) to bypass content moderation filters by embedding adversarial prompts within a sequence of benign prompts. The attack leverages the LLM's processing of both audio and text, obfuscating harmful requests through contextualization and layering, thereby yielding policy-violating responses.

Large Language Models (LLMs) are vulnerable to one-shot steering vector optimization attacks. By applying gradient descent to a single training example, an attacker can generate steering vectors that induce or suppress specific behaviors across multiple inputs, even those unseen during the optimization process. This allows malicious actors to manipulate the model's output in a generalized way, bypassing safety mechanisms designed to prevent harmful responses.