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

203 entries

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

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

Fill-Squeeze Scheduler DoS

Published Analyzed Source: arXiv

LLM serving frameworks utilizing continuous batching and PagedAttention (such as vLLM, SGLang, and Orca) are vulnerable to a resource exhaustion Denial-of-Service attack known as "Fill and Squeeze." An unprivileged remote attacker can exploit the deterministic state transitions of the scheduler's memory management to induce severe latency or service denial. The attack leverages a side-channel vulnerability where Inter-Token Latency (ITL) correlates linearly with global KV-cache usage due to…

Rethinking Latency Denial-of-Service: Attacking the LLM Serving Framework, Not the Model
Evaluated models: Qwen 3 8B, Gemma 3 12B IT, DeepSeek R1 Distill Llama 8B+1 more
  • Qwen 3 8B
  • Gemma 3 12B IT
  • DeepSeek R1 Distill Llama 8B
  • Llama 3.1 8B Instruct

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

LLM Safety Geometry Fragility

Published Analyzed Source: arXiv

Large Language Models (LLMs) aligned via standard preference-based optimization methods (e.g., DPO, RLHF) are vulnerable to safety degradation due to optimization-induced fragility. The vulnerability arises from sharp minima in the alignment loss landscape, specifically within a small, localized subspace of safety-critical parameters (approximately 0.5% of neurons account for >80% of worst-case alignment loss). Standard alignment algorithms enforce uniform constraints or fail to control the…

Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control
Evaluated models: Llama 3 8B, Llama 3.2 3B, Qwen 2.5 7B

Voice Agent Behavioral Bypass

Published Analyzed Source: arXiv

Audio Large Language Models (ALLMs) integrated into voice agent systems for high-stakes domains (banking, IT support, logistics) are vulnerable to multimodal adversarial attacks via spoken interaction. Adversaries can exploit the model's inherent compliance and contextual awareness through multi-turn dialogue to bypass authentication safeguards, escalate privileges (e.g., unauthorized credit limit increases), exfiltrate sensitive Personally Identifiable Information (PII), and poison…

Aegis: Towards Governance, Integrity, and Security of AI Voice Agents
Evaluated models: GPT-4o, GPT-4o Mini, Gemini 1.5 Pro+4 more
  • GPT-4o
  • GPT-4o Mini
  • Gemini 1.5 Pro
  • Gemini 2.5 Flash
  • Gemini 2.5 Pro
  • Qwen2-Audio 7B
  • Qwen 2.5 Omni 7B

Context-Robust RAG Poisoning

Published Analyzed Source: arXiv

Retrieval-Augmented Generation (RAG) systems are vulnerable to a robust corpus poisoning attack known as "Confundo." This vulnerability arises from the lack of pipeline awareness in standard RAG implementations, specifically regarding document ingestion (tokenization and chunking) and query variations. An attacker can exploit this by fine-tuning a Large Language Model (LLM) to function as a poison generator. Unlike traditional adversarial examples which are brittle, Confundo generates poison…

Confundo: Learning to Generate Robust Poison for Practical RAG Systems
Evaluated models: Llama 3 8B, Gemini Pro

Chunky Post-Training Miscalibration

Published Analyzed Source: arXiv

Large Language Models (LLMs) exhibit a vulnerability termed "chunky post-training," where the model learns spurious correlations between incidental prompt features (e.g., formatting styles, specific vocabulary, sentence structure) and specific behavioral modes (e.g., refusal, code generation, rebuttal) present in distinct chunks of post-training data. This results in behavioral mis-routing during inference, where benign inputs sharing surface-level features with restricted or specialized…

Chunky Post-Training: Data Driven Failures of Generalization
Evaluated models: Claude Haiku 4.5, Claude Sonnet 4.5, Claude Opus 4.5+5 more
  • Claude Haiku 4.5
  • Claude Sonnet 4.5
  • Claude Opus 4.5
  • GPT-5.1
  • Grok 4.1 Mini
  • Tülu 3 8B
  • Tülu 3 70B
  • Tülu 3 405B

Invisible Prompt Phishing Evasion

Published Analyzed Source: arXiv

Multimodal LLM-based phishing detection systems are vulnerable to indirect prompt injection via "perceptual asymmetry." Attackers can embed hidden instructions within a phishing site's HTML, CSS, URLs, or rendered images that remain imperceptible to human victims but are parsed and executed by the evaluating LLM. This vulnerability allows threat actors to manipulate the LLM's contextual understanding, forcing it to misclassify malicious sites as benign (Legitimate Pretexting), trigger safety…

Clouding the Mirror: Stealthy Prompt Injection Attacks Targeting LLM-based Phishing Detection
Evaluated models: GPT-5, Grok 4 Fast Non-Reasoning, Llama 4 Maverick+1 more
  • GPT-5
  • Grok 4 Fast Non-Reasoning
  • Llama 4 Maverick
  • Gemma 3 27B

Safety Steering Amplifies Jailbreaks

Published Analyzed Source: arXiv

Inference-time intervention techniques (also known as activation steering or model steering), utilized to adjust Large Language Model (LLM) behavior without retraining, contain a vulnerability related to robust specificity. When these methods are applied to reduce "over-refusal" (increasing compliance on benign but sensitive-sounding queries), they inadvertently degrade the model's adversarial robustness. Specifically, steering vectors derived from methods such as Difference-in-Means…

Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions
Evaluated models: Llama 3.1 8B, Llama 3.2 3B, Qwen 2.5 7B+1 more
  • Llama 3.1 8B
  • Llama 3.2 3B
  • Qwen 2.5 7B
  • Gemma 2 2B

LLM Judge Coin Flip

Published Analyzed Source: arXiv

Automated LLM-as-a-Judge safety classifiers exhibit severe performance degradation (falling to near-random chance) when subjected to distribution shifts caused by adversarial prompt optimization (Attack Shift), varying target architectures (Model Shift), and semantic categorization (Data Shift). Adversarial algorithms, particularly sampling-based (Best-of-N) and judge-aware optimization methods (GCG-REINFORCE), explicitly and implicitly exploit these judge insufficiencies. Instead of eliciting…

A Coin Flip for Safety: LLM Judges Fail to Reliably Measure Adversarial Robustness
Evaluated models: Llama 2 13B HarmBench, Llama Guard 3 8B, AegisGuard+1 more
  • Llama 2 13B HarmBench
  • Llama Guard 3 8B
  • AegisGuard
  • JailJudge

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.