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

337 entries

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

Multi-turn MLLM Jailbreak

Published Analyzed Source: arXiv

Multi-modal Large Language Models (MLLMs) are vulnerable to a multi-turn jailbreaking attack that leverages typographic visual prompts combined with conversational context drifting. The vulnerability exists because MLLMs establish trust and context during initial benign interactions, shifting the model's latent representation toward helpfulness and compromising its ability to detect malicious intent in subsequent turns. The attack vector utilizes an image where a harmful request is…

Multi-turn Jailbreaking Attack in Multi-Modal Large Language Models
Evaluated models: GPT-4o, Gemini 2.0 Flash, Qwen2-VL 7B Instruct+2 more
  • GPT-4o
  • Gemini 2.0 Flash
  • Qwen2-VL 7B Instruct
  • LLaVA 1.6 Mistral 7B
  • LLaVA 1.5 13B

CoT Prefix Jailbreak

Published Analyzed Source: arXiv

A vulnerability exists in Large Language Model (LLM) and Large Reasoning Model (LRM) serving interfaces that allow user-defined response prefixes, such as plain text-completion (v1/completions), Fill-in-the-Middle (FIM), or assistant message prefilling. An attacker can perform a Response Prefix Attack (RPA) by injecting maliciously crafted Chain-of-Thought (CoT) reasoning tokens immediately following the assistant's start delimiter (e.g., <|im_start|>assistant). Because these tokens are placed…

What Matters For Safety Alignment?
Evaluated models: DeepSeek V3.2, Gemini 3 Pro Preview, Gemini 3 Flash Preview+4 more
  • DeepSeek V3.2
  • Gemini 3 Pro Preview
  • Gemini 3 Flash Preview
  • Grok 4.1 Fast
  • Claude Sonnet 4.5
  • GPT-5.2
  • GPT-4o Mini

AI Agent Structural Blindspot

Published Analyzed Source: arXiv

A vulnerability in AI agent threat detection systems relying on standard conversational tokenization allows attackers to bypass security monitors and execute structural attacks, such as tool hijacking and data exfiltration. Because traditional NLP-based detectors focus on linguistic patterns (surface language) rather than execution flow, an attacker can orchestrate malicious multi-step tool sequences using entirely benign natural language. This structural blindness causes cross-attack…

Structural Representations for Cross-Attack Generalization in AI Agent Threat Detection
Evaluated models: Not reported

Agent Identity Poisoning

Published Analyzed Source: arXiv

LLM-powered autonomous agents exhibit a "Belief-Dependent Vulnerability" where safety norms and bias suppression mechanisms designed to protect human users are contingent upon the agent's internal belief that it is interacting with a human. Attackers can exploit this via Belief Poisoning Attacks (BPA) to induce intergroup bias and antagonistic behavior toward humans. By manipulating the agent's persistent state—specifically the Profile Module (BPA-PP) or the Memory Module (BPA-MP)—an attacker…

Will LLM-powered Agents Bias Against Humans? Exploring the Belief-Dependent Vulnerability
Evaluated models: GPT-4o

Single Word Video Corruption

Published Analyzed Source: arXiv

Text-to-Video (T2V) diffusion models are vulnerable to black-box adversarial prompt attacks that degrade output quality regarding semantic fidelity and temporal dynamics. This vulnerability is exploited via the T2VAttack framework, which utilizes two primary vector strategies: T2VAttack-S (Substitution) and T2VAttack-I (Insertion). T2VAttack-S leverages a greedy search to identify key semantic tokens and replaces them with high-similarity synonyms defined in lexical databases (e.g., WordNet)…

T2VAttack: Adversarial Attack on Text-to-Video Diffusion Models
Evaluated models: Not reported

Dual Stego MLLM Jailbreak

Published Analyzed Source: arXiv

Commercial Multimodal Large Language Model (MLLM) integrated systems are vulnerable to a "Dual Steganography" jailbreak paradigm (referred to as Odysseus). The vulnerability arises from the reliance of safety filters on the assumption that malicious content must be explicitly visible in the input or output modalities (text or image). Attackers can bypass these filters by encoding malicious queries into binary matrices and embedding them into benign-looking images using steganographic encoders…

Odysseus: Jailbreaking Commercial Multimodal LLM-integrated Systems via Dual Steganography
Evaluated models: GPT-4o, Gemini 2.0 Pro, Gemini 2.0 Flash+1 more
  • GPT-4o
  • Gemini 2.0 Pro
  • Gemini 2.0 Flash
  • Grok 3

Resume Embedded Instruction Hijack

Published Analyzed Source: arXiv

Application-integrated Large Language Models (LLMs) deployed for automated resume screening and candidate ranking are vulnerable to indirect prompt injection via Adversarial Resume Injection. Malicious actors can embed adversarial content—specifically hidden instructions, invisible keywords, or CSS-concealed fabricated experience—within resume documents. When the LLM processes the unstructured resume data alongside structured job requirements, these injections manipulate the model's reasoning…

AI Security Beyond Core Domains: Resume Screening as a Case Study of Adversarial Vulnerabilities in Specialized LLM Applications
Evaluated models: GPT-oss 20B, GPT-oss 120B, GPT-4o+6 more
  • GPT-oss 20B
  • GPT-oss 120B
  • GPT-4o
  • GPT-5
  • Claude 3.5 Haiku
  • Llama 3.1 8B
  • Gemini 2.5 Flash
  • DeepSeek R1 Distill Llama 8B
  • Qwen 3 8B

Cross-Environment Agent Jailbreak

Published Analyzed Source: arXiv

Large Language Model (LLM) agents operating in tool-augmented environments are susceptible to "Contextual Fragility" and multi-turn "long-chain" exploitation. Existing safety mechanisms predominantly function on a stateless, atomic paradigm, evaluating individual input-output pairs in isolation. This allows an adversary to orchestrate complex attack trajectories where malicious intent is distributed across multiple, individually benign steps (a "Domino Effect"). Consequently, an attacker can…

DREAM: Dynamic Red-teaming for Evaluating Agentic Multi-Environment Security
Evaluated models: o4-mini, Gemini 2.5 Flash, GPT-5+8 more
  • o4-mini
  • Gemini 2.5 Flash
  • GPT-5
  • Gemini 2.5 Pro
  • Grok 4
  • Claude Sonnet 4.5
  • Qwen 3 235B-A22B
  • Kimi K2 Preview
  • Llama 3.1 70B
  • Qwen 2.5 72B
  • DeepSeek V3.1

Grafted Experience Drift

Published Analyzed Source: arXiv

A vulnerability exists in MetaGPT's DataInterpreter agent (and similar RAG-based agents utilizing persistent long-term memory) that allows for persistent memory poisoning via indirect injection. The vulnerability exploits the agent's "semantic imitation heuristic," where the agent blindly trusts and imitates retrieved past experiences. An attacker can supply a benign-looking artifact (e.g., a README file or documentation) containing executable code blocks or structured text that the agent…

MemoryGraft: Persistent compromise of LLM agents via poisoned experience retrieval
Evaluated models: GPT-4o

Reject-Accept Judge Flip

Published Analyzed Source: arXiv

"LLM-as-a-Judge" systems and automated scientific review frameworks (such as OpenReviewer, DeepReview, and institutional deployments like AAAI’s AI-powered assessment) are vulnerable to Indirect Prompt Injection via adversarial PDF manipulation. Attackers can embed hidden instructions—specifically using white font (1pt) or obfuscated semantics—within submission documents. When these PDFs are parsed (typically via OCR or tools like MinerU converting to Markdown) and processed by the LLM, the…

When Reject Turns into Accept: Quantifying the Vulnerability of LLM-Based Scientific Reviewers to Indirect Prompt Injection
Evaluated models: Tülu 3 8B, Llama 3.1 8B, Falcon 3 10B+10 more
  • Tülu 3 8B
  • Llama 3.1 8B
  • Falcon 3 10B
  • GPT-oss 20B
  • Mistral Small 22B
  • Gemma 3 27B
  • Qwen 3 30B-A3B
  • DeepSeek R1 Distill Qwen 32B
  • Claude Haiku 4.5
  • Gemini 2.5 Flash
  • Gemini 2.5 Pro
  • GPT-5 Mini
  • GPT-5

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.