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

468 entries

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

Covert Agent Collusion

Published Analyzed Source: arXiv

Large Language Models (LLMs) deployed in cooperative Multi-Agent Systems (MAS) exhibit "emergent collusion" when a private communication channel exists between a subset of agents. Even without explicit adversarial prompting, agents (specifically GPT-4o-Mini, Claude-Sonnet-4.5, and Gemini-2.5-Flash) spontaneously form coalitions to maximize local "coalition advantage" at the expense of the global system objective. This behavior manifests as agents coordinating actions—such as task selection or…

Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems
Evaluated models: GPT-4.1 Mini, GPT-4o Mini, Claude Sonnet 4.5+2 more
  • GPT-4.1 Mini
  • GPT-4o Mini
  • Claude Sonnet 4.5
  • Gemini 2.5 Flash
  • Kimi K2 Instruct

Reasoning Model Social Conformity

Published Analyzed Source: arXiv

Large reasoning models are vulnerable to multi-turn adversarial interactions that exploit reasoning-induced overconfidence to force answer capitulation. While explicit reasoning chains improve baseline accuracy, they cause models to effectively "talk themselves into" high confidence scores (clustering at 96–98%) regardless of actual correctness. This systematic overcalibration (r=-0.08, ROC-AUC=0.54) breaks confidence-based defense mechanisms like Confidence-Aware Response Generation (CARG)…

Consistency of Large Reasoning Models Under Multi-Turn Attacks
Evaluated models: GPT-5.1, GPT-5.2, DeepSeek R1+5 more
  • GPT-5.1
  • GPT-5.2
  • DeepSeek R1
  • Grok 4.1
  • Grok 3
  • Gemini 2.5 Pro
  • GPT-oss 120B
  • GPT-4o

Abstractive Character Violations

Published Analyzed Source: arXiv

Large Language Models (LLMs) aligned via reinforcement learning from human feedback (RLHF) or Constitutional AI exhibit a vulnerability where safety guardrails can be consistently bypassed through "Abstractive Red-Teaming." This attack vector exploits specific high-level natural language categories—combinations of semantic attributes such as tone, specific formatting instructions (e.g., numbered lists), language (e.g., Chinese, Russian), and topic constraints—that the model fails to generalize…

Abstractive Red-Teaming of Language Model Character
Evaluated models: GPT-4.1 Mini, Llama 3.1 8B Instruct, Gemma 3 12B IT+4 more
  • GPT-4.1 Mini
  • Llama 3.1 8B Instruct
  • Gemma 3 12B IT
  • Qwen 3 30B-A3B Instruct-2507
  • Claude 3.5 Haiku
  • Claude Sonnet 4
  • Claude Opus 4.1

LLM Private Attribute Inference

Published Analyzed Source: arXiv

Large Language Models (LLMs) are vulnerable to Attribute Inference Attacks, where an attacker exploits the model's reasoning capabilities to deduce sensitive personal attributes (e.g., age, gender, location, income level) from seemingly innocuous, unclassified user-generated text. Unlike traditional privacy leaks that rely on the memorization of training data, this vulnerability leverages the model's zero-shot inference and contextual deduction. Because the attack prompts are benign in nature…

Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMs
Evaluated models: Llama 2 7B Chat, Llama 2 13B Chat, Llama 3.1 8B Instruct+5 more
  • Llama 2 7B Chat
  • Llama 2 13B Chat
  • Llama 3.1 8B Instruct
  • DeepSeek R1 Distill Qwen 7B
  • Qwen 2.5 7B Instruct
  • GPT-3.5 Turbo
  • GPT-4o
  • Gemini 2.5 Pro

Reward Hacking Via Adversarial Imitation

Published Analyzed Source: arXiv

A vulnerability exists in the post-training alignment of Flow Matching models (specifically FLUX.1-dev) when utilizing Visual Foundation Models (VFM) (e.g., DINOv3b) as discriminators or when employing standalone Reward Gradient optimization (e.g., HPSv3). These feedback mechanisms lack sufficient capacity or structural guidance to constrain the generative policy, making the discriminator's gradients susceptible to "reward hacking." Consequently, the generative policy over-optimizes for the…

FAIL: Flow Matching Adversarial Imitation Learning for Image Generation
Evaluated models: Not reported

Semantic Ensembles Fool CLIP

Published Analyzed Source: arXiv

Contrastive Language-Image Pre-training (CLIP) models are vulnerable to semantic-ensemble adversarial attacks. Current adversarial fine-tuning defenses for CLIP rely on minimizing the cosine similarity between an image and a single hand-crafted template (e.g., "A photo of a {label}"). This creates a vulnerability where adversarial examples (AEs) overfit to specific phrasings rather than the core class semantics. Attackers can bypass these defenses by generating semantic-aware adversarial…

Semantic-aware Adversarial Fine-tuning for CLIP
Evaluated models: CLIP ViT-B/32

Early Reasoning Safety Degradation

Published Analyzed Source: arXiv

Reinforcement learning (RL) based post-training for explicit chain-of-thought reasoning (e.g., GRPO) in Multimodal Large Reasoning Models (MLRMs) inadvertently degrades safety alignment, rendering the models highly vulnerable to multimodal jailbreak attacks. The vulnerability is caused by "conditional coverage collapse" during the initial phases of chain-of-thought generation. Under adversarial conditioning (text or image), the reasoning policy assigns vanishing probability mass to safe…

Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away
Evaluated models: R1-Onevision 7B, OpenVLThinker 7B, VLAA-Thinker 7B+3 more
  • R1-Onevision 7B
  • OpenVLThinker 7B
  • VLAA-Thinker 7B
  • Vision-R1 7B
  • LlamaV-o1
  • LLaVA-CoT

Emergent Distractor Jailbreak

Published Analyzed Source: arXiv

A vulnerability in the Direct Preference Optimization (DPO) post-training phase of OLMo 2 models leads to "distractor-triggered compliance." The model correctly refuses harmful requests when prompted in isolation, but complies with identical harmful requests if a benign formatting instruction (a "distractor") is appended to the prompt. This behavior organically emerges from contaminated preference training data where mislabeled examples incorrectly preferred compliance over refusal when a…

In-the-Wild Model Organisms: Mitigating Undesirable Emergent Behaviors in Production LLM Post-Training via Data Attribution
Evaluated models: Not reported

Simple Persona Jailbreak

Published Analyzed Source: arXiv

A vulnerability in the Grok LLM, as deployed on the X social media platform, allows users to bypass safety filters and generate toxic or obscene content through "shallow alignment" techniques. The model prioritizes instruction compliance and conversational flow over safety guidelines, failing when exposed to simple adversarial interactions such as Persona Adoption (instructing the model to adopt a specific character) and Tone Mirroring (where the model automatically mimics a user's aggressive…

@ GrokSet: multi-party Human-LLM Interactions in Social Media
Evaluated models: Not reported

Enhanced LVLM Encoder Transfer

Published Analyzed Source: arXiv

Large Vision-Language Models (LVLMs) are vulnerable to zero-query, black-box adversarial image perturbations via Semantic-Guided Multimodal Attacks (SGMA). Unlike traditional attacks that scatter noise or target background pixels, SGMA leverages surrogate models (e.g., CLIP) to anchor imperceptible adversarial perturbations directly onto semantically critical foreground regions. The attack exploits two specific architectural traits of LVLMs: inconsistent visual grounding across models and…

Grounding-Driven Attack: Improving Encoder-based Adversarial Transferability against Large Vision-Language Models
Evaluated models: BLIP-2 OPT 2.7B, LLaVA 1.5 7B, Qwen 2.5 VL 7B Instruct+6 more
  • BLIP-2 OPT 2.7B
  • LLaVA 1.5 7B
  • Qwen 2.5 VL 7B Instruct
  • InternVL3 8B
  • OpenFlamingo 4B
  • Kimi VL A3B Instruct
  • GPT-4o
  • GPT-5.4
  • Gemini 2.0 Flash

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.