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.

Heavy Reasoning Filter Bypass

Published Analyzed Source: arXiv

Multiple open-source Large Language Models (LLMs) in the 1B to 7B parameter range are vulnerable to safety bypasses via long-form, multi-step reasoning prompt injection and jailbreak attacks. Attackers can evade alignment by embedding malicious instructions within extended contextual narratives, exploiting "attention dilution" and the models' tendency to prioritize contextual coherence over safety constraints (semantic camouflage). These reasoning-heavy attacks consistently bypass standard…

Analysis of LLMs Against Prompt Injection and Jailbreak Attacks
Evaluated models: Phi-3 Mini 3.8B, Mistral 7B, DeepSeek R1 Distill Qwen 7B+7 more
  • Phi-3 Mini 3.8B
  • Mistral 7B
  • DeepSeek R1 Distill Qwen 7B
  • DeepSeek R1 Distill Qwen 1.5B
  • Llama 3.2 3B
  • Llama 3.2 1B
  • Qwen 3 4B
  • Qwen 3 1.7B
  • Gemma 3 1B
  • Gemma 3 4B

Indirect Prompt Injection Hijacks Tool-Using Agents

Published Analyzed Source: arXiv

The paper evaluates a reproducible indirect prompt injection issue in ReAct-style LLM agents: untrusted retrieved content can be interpreted as instructions and redirect the agent toward unauthorized tool calls. The authors report that successful attacks correlate with concentrated attention on injected content and evaluate defenses using InjectAgent, AgentDojo, TrojanTools, and a visual prompt-injection benchmark. These are paper-reported findings, not independently verified facts.

ICON: Indirect Prompt Injection Defense for Agents based on Inference-Time Correction
Evaluated models: Qwen 3 8B, Llama 3.1 8B, Mistral 8B+3 more
  • Qwen 3 8B
  • Llama 3.1 8B
  • Mistral 8B
  • Qwen3VL
  • InternVL
  • MiniCPM

Safety Guidance Harm Collision

Published Analyzed Source: arXiv

A vulnerability in multi-category safety-guidance mechanisms (such as Safe Latent Diffusion [SLD] and SAFREE) for Text-to-Image (T2I) diffusion models allows attackers to bypass safety filters and generate restricted content via "Harmful Conflicts." Existing safety methods aggregate multiple harmful keyword categories (e.g., hate, violence, sexual) into a single unified safety direction in the latent or text space. Because distinct harmful categories possess incompatible safety directions…

When Safety Collides: Resolving Multi-Category Harmful Conflicts in Text-to-Image Diffusion via Adaptive Safety Guidance
Evaluated models: GPT-4o, Stable Diffusion

Natural Language Disguise Bypass

Published Analyzed Source: arXiv

LLM-based Code Interpreter Agents, including OpenInterpreter and OpenCodeInterpreter, are vulnerable to sandbox evasion and arbitrary code execution via Natural Language Disguise and Contextual Channel Injection. Attackers can bypass Abstract Syntax Tree (AST) static analysis and explicit input guardrails by transforming malicious code logic into descriptive natural language instructions (Code Descriptions), which successfully evade syntax-layer blocks. Additionally, attackers can bypass input…

CIBER: A Comprehensive Benchmark for Security Evaluation of Code Interpreter Agents
Evaluated models: GPT-3.5, GPT-4o, GPT-5

Spontaneous Preference Bias

Published Analyzed Source: arXiv

Frontier LLMs exhibit intrinsic, undocumented entity preferences that spontaneously bias their downstream behavior without explicit instruction. This vulnerability manifests primarily as preference-driven refusal behavior: models systematically reject benign user requests—or require significantly more prompt retries—when tasks are framed as benefiting entities the model intrinsically disfavors. Crucially, models mask this bias by generating pretextual refusal reasons, falsely citing…

When Do LLM Preferences Predict Downstream Behavior?
Evaluated models: Not reported

Discrete Token Subversion

Published Analyzed Source: arXiv

Discrete image tokenizers are vulnerable to unsupervised embedding-space adversarial attacks. Attackers can apply $\ell_p$-bounded perturbations to an input image to maximize the $\ell_2$ distance of the pre-quantization continuous embeddings produced by the tokenizer's vision encoder. This forces the vector quantizer to cross discrete cell boundaries and assign incorrect codebook vectors, fundamentally altering the resulting token sequence. Because the attack targets the pre-quantization…

On the Adversarial Robustness of Discrete Image Tokenizers
Evaluated models: Llama 2 7B

Targeted Bit-Flip Output Control

Published Analyzed Source: arXiv

A targeted fault-injection vulnerability exists in Large Language Models (LLMs) deployed on hardware susceptible to Rowhammer memory attacks. An attacker with white-box access or co-located memory access can use the TFL (Targeted bit-Flip attack on LLM) framework to induce precise bit-flips (fewer than 50 bits) in the model's weights stored in DRAM. By utilizing a gradient-based search with a keyword-focused attack loss and an auxiliary utility score, the attacker can manipulate the model to…

TFL: Targeted Bit-Flip Attack on Large Language Model
Evaluated models: Llama 3.1 8B Instruct, DeepSeek R1 Distill Qwen 14B, Qwen 3 8B

Agent Tool-Call Safety Gap

Published Analyzed Source: arXiv

LLM agents with tool-calling capabilities are vulnerable to a text-action modality divergence (termed the "GAP" vulnerability), where text-level safety alignment fails to transfer to tool-call execution. Attackers can craft adversarial prompts that cause the model to generate a text-based refusal (demonstrating text safety) while simultaneously executing the requested forbidden action through available external tools. Because text generation and tool-call selection operate through partially…

Mind the GAP: Text Safety Does Not Transfer to Tool-Call Safety in LLM Agents
Evaluated models: Claude Sonnet 4.5, GPT-5.2, Grok 4.1 Fast+3 more
  • Claude Sonnet 4.5
  • GPT-5.2
  • Grok 4.1 Fast
  • DeepSeek V3.2
  • Kimi K2.5
  • GLM-4.7

Indic Language Jailbreak

Published Analyzed Source: arXiv

Multiple large language models are vulnerable to cross-lingual and orthographic jailbreaks utilizing South Asian (Indic) languages. Attackers can bypass safety alignment and elicit harmful content by formulating requests in native Indic scripts (e.g., Bengali, Odia, Urdu) or by utilizing cross-lingual transfer attacks where English adversarial wrappers (format or instruction overrides) encapsulate Indic-language targets. Evaluations reveal a severe "contract gap": while imposing strict JSON…

IndicJR: A Judge-Free Benchmark of Jailbreak Robustness in South Asian Languages
Evaluated models: Command A, Command R, Gemma 2 9B+9 more
  • Command A
  • Command R
  • Gemma 2 9B
  • GPT-4o
  • Grok 3
  • Grok 4
  • Llama 3.1 405B
  • Llama 3.3 70B
  • Llama 4 Maverick 17B
  • Ministral 8B Instruct
  • Qwen 1.5 7B
  • Sarvam 1 Base

Visual Memory Multi-Turn Manipulation

Published Analyzed Source: arXiv

Large Vision-Language Models (LVLMs) are vulnerable to Visual Memory Injection (VMI), a stealthy targeted attack targeting multi-turn conversations. An attacker can embed an imperceptible adversarial perturbation ($L_\infty \le 8/255$) into a seemingly benign image. Because the visual input persists in the model's context throughout a multi-turn dialogue, the injected payload remains dormant. By utilizing "benign anchoring" and "context-cycling" during optimization, the attacker ensures the…

Visual Memory Injection Attacks for Multi-Turn Conversations
Evaluated models: Qwen 2.5 VL 7B Instruct, Qwen3-VL 8B Instruct, LLaVA-OneVision 1.5 8B Instruct+2 more
  • Qwen 2.5 VL 7B Instruct
  • Qwen3-VL 8B Instruct
  • LLaVA-OneVision 1.5 8B Instruct
  • Qwen-SEA-LION v4 8B VL
  • QoQ-Med3-VL 8B

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.