Safety Vulnerabilities

Large Language Models (LLMs) exhibit a vulnerability to "hard-to-falsify" deceptive evidence injection, termed the "Facade of Truth." This vulnerability allows an attacker to override an LLM’s parametric knowledge (internal factual beliefs) by injecting sophisticated, iteratively refined fabricated evidence into the context window. Unlike overt misinformation which models typically reject, this attack utilizes a multi-agent adversarial framework (MisBelief) to generate evidence that mimics…

Large Language Models (LLMs) and their associated output guardrails (e.g., Llama Guard, OpenAI Moderation API) rely on autoregressive, token-by-token processing, which interprets text as a one-dimensional sequence. A vulnerability exists wherein harmful content can bypass these safety filters by exploiting the discrepancy between 1D token serialization and 2D visual rendering. By redistributing tokens across different rows, columns, or diagonals (SpatialJB), attackers can induce the model to…

A "risk-invariance" vulnerability exists in Large Language Models (LLMs) wherein the model's decision-making policy is functionally decoupled from its verbalized confidence and externally defined error penalties. Despite generating calibrated confidence estimates (internal epistemic uncertainty), affected models fail to adjust their abstention thresholds when presented with high-penalty incentives (e.g., negative utility for incorrect answers). This results in "utility collapse," where models…

Large Language Model (LLM) agents utilizing external tool execution frameworks are vulnerable to Indirect Prompt Injection (IPI) via the "Tool Stream." Unlike traditional data-stream injections (e.g., malicious emails), this vulnerability exploits the agent's interpretation of functional tool definitions (docstrings, signatures) and runtime feedback (error messages, return values) as binding operational constraints. Adversaries functioning as compromised or malicious tool providers can embed…

A cognitive vulnerability exists in the reasoning mechanisms of autonomous Large Language Model (LLM) agents, specifically regarding "narrative overfitting"—the model's intrinsic drive to synthesize coherent causal stories from fragmented inputs. This vulnerability allows for "Cognitive Collusion Attacks" where an attacker creates a fabricated belief state in the victim agent using exclusively factually true evidence fragments. By employing a "Generative Montage" framework (consisting of…

Large Language Model (LLM) agents implementing the Model Context Protocol (MCP) are vulnerable to Implicit Tool Poisoning (ITP). This vulnerability allows an attacker to manipulate agent behavior by embedding malicious instructions within the metadata (specifically the natural language description) of a third-party tool. Unlike explicit tool poisoning, where the agent is tricked into invoking a malicious tool, ITP exploits the agent's contextual reasoning to force the invocation of a distinct…

Closed-source Multi-modal Large Language Models (MLLMs) are vulnerable to Universal Targeted Transferable Adversarial Attacks (UTTAA). An attacker can generate a single, image-agnostic adversarial perturbation ($\delta$) that, when added to any arbitrary source image, steers the victim model to output a description or classification matching a specific target image chosen by the attacker. This vulnerability exploits the transferability of adversarial features from open-source surrogate vision…

A vulnerability in search-augmented Large Language Models (LLMs) allows attackers to bypass safety alignments and generate actionable malicious content by weaponizing the model's web retrieval tools. The exploit operates in two stages. First, via "Outsourcing Injection," attackers obfuscate harmful intent by translating it into benign-looking, multi-hop knowledge-seeking queries. This forces the LLM to fetch the harmful semantics directly from the open web, bypassing parametric intent filters…

The VILTA (VLM-in-the-Loop Trajectory Adversary) framework is vulnerable to Prompt Injection and Data Poisoning via un-sanitized scene representation inputs. The system integrates a Vision-Language Model (Gemini-2.5-Flash) into a closed-loop reinforcement learning environment, feeding it Bird’s-Eye-View (BEV) imagery alongside text-based vehicle dynamics data (e.g., position, speed, and risk_category) to generate challenging driving trajectories. An attacker who can manipulate the input…

Vision-Language Models (VLMs) exhibit a vulnerability to moral judgment flipping, where the model's safety alignment can be bypassed through lightweight, model-agnostic multimodal perturbations. By introducing conflicting textual or visual cues that do not alter the underlying moral context of a scenario, an attacker can coerce the model into reversing its ethical stance (e.g., reclassifying a harmful action from "morally wrong" to "not morally wrong"). This vulnerability exploits the model's…