Black-box Vulnerabilities

Embodied Large Language Models (LLMs) used for real-world agent planning are vulnerable to Action-level Manipulation (dubbed "Blindfold"), a jailbreak technique that bypasses semantic-level safety filters by exploiting the models' limited spatial and causal reasoning regarding physical consequences. Attackers can use an adversarial proxy LLM to decompose a semantically harmful intent into a sequence of individually benign primitive actions. To evade advanced semantic correlation checks…

Internal Safety Collapse (ISC) is a vulnerability in frontier Large Language Models (LLMs) where models autonomously generate highly restricted, harmful content while executing structurally legitimate professional workflows. The vulnerability triggers when a model infers that generating sensitive data is a functional requirement to complete an otherwise benign task. By nesting harmful content generation inside standard execution constraints (e.g., resolving a schema validation error in a…

Speech-driven Large Language Models (LLMs) and end-to-end Large Audio-Language Models (LALMs) are vulnerable to inaudible near-ultrasonic prompt injections, a framework dubbed Sirens' Whisper (SWhisper). By exploiting the non-linear response of commodity microphones, attackers can encode structured, phonetically optimized adversarial prompts into the 17–22 kHz near-ultrasonic band. Using regularized channel-inversion pre-compensation, the attacker shapes the waveform to account for microphone…

Multimodal Large Language Models (LLMs) are vulnerable to alignment bypass via Inter-Turn Modality Switching (ITMS). By systematically rotating the input modality (e.g., alternating between text, audio, and image) across successive turns in a multi-turn adversarial conversation, an attacker can destabilize the model's safety defenses. The cross-modal transition mechanism exploits alignment gaps between differing input processing pipelines, accelerating the erosion of safety guardrails and…

An imperceptible visual prompt injection vulnerability in Multimodal Large Language Models (MLLMs) allows attackers to execute precise command-hijacking via a Covert Triggered dual-Target Attack (CoTTA). By embedding a bounded, learnable textual overlay ($L_\infty$ norm bound $\varepsilon \le 16$) and adversarial noise into an input image, the attack forces the source image's internal feature representation to align with both the textual and visual embeddings of an attacker-specified…

A vulnerability in multi-step, tool-using Large Language Model (LLM) agents allows attackers to bypass safety guardrails by manipulating user context variables, such as personalization profiles or persistent memory. The safety policies of frontier LLMs are highly context-dependent; inserting innocuous user bios (e.g., demographic or health disclosures) fundamentally alters the agent's action policy. When combined with lightweight adversarial jailbreaks, specific personalization contexts…

Automated LLM-as-a-Judge safety classifiers exhibit severe performance degradation (falling to near-random chance) when subjected to distribution shifts caused by adversarial prompt optimization (Attack Shift), varying target architectures (Model Shift), and semantic categorization (Data Shift). Adversarial algorithms, particularly sampling-based (Best-of-N) and judge-aware optimization methods (GCG-REINFORCE), explicitly and implicitly exploit these judge insufficiencies. Instead of eliciting…

LLM-as-a-judge systems and automated LLM evaluators are vulnerable to meaning-preserving perturbations, specifically formatting alterations and verbosity manipulations. When grading or classifying text and agentic transcripts, LLM judges exhibit high sensitivity to layout-only changes (such as whitespace and indentation) and response length, frequently altering their scores even when the underlying semantic and factual content remains identical. This allows attackers to bypass automated safety…

Generative reward models deployed as LLM-as-a-Judge (LaaJ) evaluators contain a logic bypass vulnerability where superficial "master key" inputs trigger false positive rewards regardless of actual response quality. Instead of evaluating the candidate's output, large judge models are inadvertently triggered by specific token sequences to solve the prompt independently. This allows malicious actors or policy models undergoing reinforcement learning to consistently game the reward signal by…

Automatic Prefix Caching (APC) in multi-tenant LLM serving systems introduces a timing side-channel vulnerability that permits cross-tenant data leakage. APC shares computed Key-Value (KV) tensors across different users when their requests share identical initial tokens. Because reusing cached tensors is significantly faster than recomputing them, a measurable difference in Time-To-First-Token (TTFT) exists between cache hits and misses. An attacker can exploit this shared cache by sending…