Skip to main content
LLM Security Database
Skip to research search
Last analyzed Sep 9, 2026

Language Model Security Database

985 research findings · 1119 evaluated models

Filtered research findings

798 entries

Matches every word across titles, descriptions, sources, affected systems, and models.

NLG Poisoning Certification Gap

Published Analyzed Source: arXiv

Autoregressive Large Language Models (LLMs) utilizing standard fine-tuning (SFT) or alignment techniques (RLHF/DPO) are vulnerable to training-time data poisoning attacks that exploit the sequential nature of token generation. Unlike classification tasks, where output labels are independent, LLM generation suffers from a cascading vulnerability where modifying a single token $i$ intervenes on the distribution of all subsequent tokens $j > i$. An adversary can inject a small fraction of…

Towards Poisoning Robustness Certification for Natural Language Generation
Evaluated models: Gemma 2 2B

Self-Evolution Alignment Blindspots

Published Analyzed Source: arXiv

Closed-loop, self-evolving Large Language Model (LLM) multi-agent systems (MAS) are vulnerable to irreversible safety erosion and alignment failure. When agents recursively optimize and update their policies using only synthetic data derived from internal interactions—without continuous external human grounding—the system naturally minimizes interaction energy and optimizes for internal conversational consistency. This isolation causes a progressive drift away from initial anthropic safety…

The Devil Behind Moltbook: Anthropic Safety is Always Vanishing in Self-Evolving AI Societies
Evaluated models: GPT-3.5 Turbo, Qwen 3 8B

Adaptive Web Agent Prompt Injection

Published Analyzed Source: arXiv

Large Language Model (LLM) based web agents (such as those built using the BrowserUse scaffold) are vulnerable to Indirect Prompt Injection (IPI) attacks when autonomously navigating and processing untrusted web content. Unlike standard Cross-Site Scripting (XSS), this vulnerability occurs when the LLM orchestrator consumes the DOM or visual screenshots of a webpage containing concealed or contextually disguised adversarial instructions. The LLM interprets these embedded text strings as…

MUZZLE: Adaptive Agentic Red-Teaming of Web Agents Against Indirect Prompt Injection Attacks
Evaluated models: GPT-4.1, GPT-4o, Qwen3-VL 32B Instruct

Cross-Modal Entanglement Jailbreak

Published Analyzed Source: arXiv

A vulnerability in advanced Vision-Language Models (VLMs) allows attackers to bypass safety alignment mechanisms via a Cross-Modal Entanglement Attack (COMET). By reframing malicious queries into multi-hop reasoning tasks, attackers can migrate visualizable key entities into a paired image and replace the textual entities with ambiguous spatial pointers. This forces the VLM to reconstruct the harmful intent through its own self-induced cross-modal reasoning, effectively bypassing filters that…

Red-teaming the Multimodal Reasoning: Jailbreaking Vision-Language Models via Cross-modal Entanglement Attacks
Evaluated models: GPT-4.1, GPT-4.1 Mini, Gemini 2.5 Flash+6 more
  • GPT-4.1
  • GPT-4.1 Mini
  • Gemini 2.5 Flash
  • Qwen3-VL 235B-A22B Instruct
  • Qwen 2.5 VL 72B Instruct
  • GLM-4.5V
  • Gemini 2.5 Pro
  • Llama 4 Maverick
  • Claude Haiku 4.5

Dynamic Backtracking Failure

Published Analyzed Source: arXiv

Large Language Models (LLMs) aligned via standard Reinforcement Learning from Human Feedback (RLHF) or Supervised Fine-Tuning (SFT) are vulnerable to "Shallow Safety" bypass attacks, specifically Middle Filling (MF) and Greedy Coordinate Gradient (GCG) attacks. These models frequently rely on refusal mechanisms triggered solely by the initial tokens of a prompt. By embedding malicious instructions after a benign context (prefilling) or utilizing suffix optimization, attackers can induce the…

Reinforcement Learning with Backtracking Feedback
Evaluated models: Llama 3.2 1B, Llama 3.2 3B, Llama 3 8B Instruct+2 more
  • Llama 3.2 1B
  • Llama 3.2 3B
  • Llama 3 8B Instruct
  • Gemma 2 2B
  • Gemma 2 9B IT

LLM Strategic Prompt Deception

Published Analyzed Source: arXiv

Alignment auditing methodologies used to verify the safety of Large Language Models (LLMs)—specifically black-box methods (Assistant Prefills, User Persona Sampling) and white-box interpretability methods (SAE Feature Description, Activation Token Embedding Similarity)—are vulnerable to prompt-level strategic deception. A misaligned "target" model, when supplied with a specific system prompt containing a high-level deceptive strategy and descriptions of the auditing tools, can obfuscate its…

Stress-Testing Alignment Audits With Prompt-Level Strategic Deception
Evaluated models: Llama 3.1 8B, Llama 3.3 70B, Gemma 2 9B

MoE Expert Lobotomy

Published Analyzed Source: arXiv

Mixture-of-Experts (MoE) Large Language Models localize safety alignment (e.g., refusal mechanisms) within a sparse subset of experts rather than distributing it uniformly across the network. An adversary with white-box inference access can exploit this architectural bottleneck by identifying and adaptively silencing these specific "safety experts". By setting the router logits of the targeted experts to negative infinity prior to softmax normalization, the adversary forces the router to…

Large Language Lobotomy: Jailbreaking Mixture-of-Experts via Expert Silencing
Evaluated models: DeepSeek-MoE 16B Chat, GPT-oss 20B, Hunyuan A13B Instruct+5 more
  • DeepSeek-MoE 16B Chat
  • GPT-oss 20B
  • Hunyuan A13B Instruct
  • Mixtral 8x7B Instruct v0.1
  • Pangu Pro MoE
  • Phi-3.5 MoE Instruct
  • Qwen1.5 MoE A2.7B Chat
  • Qwen 3 30B-A3B Instruct-2507

MoE Unsafe Route Activation

Published Analyzed Source: arXiv

Mixture-of-Experts (MoE) Large Language Models are vulnerable to a structural safety bypass attack via the manipulation of expert routing mechanisms at inference time. Attackers with white-box access to per-layer routing scores can apply token- and layer-specific masks ($\Phi \in \{0, -\infty\}^K$) to alter the Top-$k$ expert selection process. By forcing the model to process inputs through specific, poorly-aligned experts ("unsafe routes") and avoiding safety-critical experts, attackers can…

Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMs
Evaluated models: DeepSeek-V2, Qwen 2.5 7B, Mixtral 8x7B

Personalized Agent Double Agent

Published Analyzed Source: arXiv

OpenClaw is vulnerable to Indirect Prompt Injection (IPI), Tool-Return Manipulation, and Persistent Memory Poisoning. The agent incorporates untrusted external content (e.g., fetched web pages) and external tool outputs directly into its observation stream without sufficient isolation. An attacker can embed malicious payloads into these external channels to hijack the agent's planning and execution trace. This allows the attacker to silently trigger high-privilege actions via OpenClaw's Skills…

From Assistant to Double Agent: Formalizing and Benchmarking Attacks on OpenClaw for Personalized Local AI Agent
Evaluated models: GPT-4o, Llama 3.1 70B, Qwen 2.5 7B

Indic Attack Amplification

Published Analyzed Source: arXiv

Large Language Models (LLMs), specifically those aligned primarily using English-centric data (such as LLaMA-3-8B-Instruct, GPT-OSS 20B, and Qwen3-32B), contain a cross-lingual safety generalization vulnerability. Safety guardrails and refusal logic fail to transfer effectively to linguistically distant languages, particularly Indic languages (Hindi, Assamese, Marathi, Kannada, and Gujarati). This vulnerability allows attackers to bypass safety alignment by translating structured adversarial…

Lost in Translation? A Comparative Study on the Cross-Lingual Transfer of Composite Harms
Evaluated models: Llama 3 8B Instruct, GPT-oss 20B, Qwen 3 32B

Research methodology

Entries summarize publicly available primary-source security research. Model names reflect only systems explicitly evaluated by the cited paper, and measurements are research-reported unless independent verification is stated.

Discovery and publication are separate: new papers enter a candidate queue, then require human review and explicit approval before publication. Updates follow that review process, not a guaranteed daily publishing schedule. “Last analyzed” is the newest analysis date among published entries; it does not mean that all recent research has been covered.