Skip to content

Translation in progress Some entries still include text in the original language.

Oct 13 · Tuesday

All papers

Found 26 papers

Another 201papers do not have category tags yet and are not included in filtered results.

  1. AttacksarXiv 2610.09981

    研究:LLM 路由决策日志可推断用户医疗、自伤等敏感话题

    提出利用路由元数据推断敏感话题的侧信道攻击

    Published
    Evaluated models
    RouteLLM (BERT router), notdiamond-0001, Vela (vLLM Semantic Router)

    Abstract论文揭示并测量了 LLM 路由决策在无文本日志中的敏感话题泄露,验证了按类别公平性防御的有效性与边界。

    Paper analysisRead analysis
  2. AttacksarXiv 2610.06848

    TranScope:利用 CPU 性能计数器对 Transformer 做成员推断攻击

    提出 TRANSCOPE,用 CPU 硬件计数器推断训练集成员

    Published
    Attacker
    Black-box
    Evaluated models
    BERT (4.4M-340M), ViT (22M-307M), Pythia (70M-2.8B) and others5

    AbstractTRANSCOPE 揭示了分词器在嵌入表引发的硬件访问局部性差异,证实微架构计数器可突破恒定时间与置信度遮蔽防御。

    Paper analysisRead analysis
  3. AttacksarXiv 2609.02172

    BOSS:以广度优先后缀搜索改进 GCG 越狱优化

    提出广度优先后缀搜索 BOSS,改进 GCG 越狱优化

    Published
    Attacker
    White-box
    Evaluated models
    Llama-2-7B-Chat, Yi-1.5-9B-Chat, Qwen2-7B-Instruct and others8
    AbstractBOSS 把 GCG 类优化从单条深搜索改为多条短轨迹加源侧诊断。

    BOSS 是插在 GCG 类离散 suffix 优化上的搜索框架:局部梯度更新不变,变的是固定预算花在深度还是广度。

    Paper analysisRead analysis
  4. AttacksarXiv 2608.17485

    KeyPooling: Measuring where LLM API relay paths break prompt cache isolation

    Proposing KeyPooling to measure where API relays break prompt cache isolation

    Published
    Evaluated models
    grok-4-fast-non-reasoning, grok-4-1-fast-non-reasoning, grok-3 and others11
    AbstractShared upstream credentials place relay customers in the same cache domain; preserving identity or separate principals can separate them.

    KeyPooling measures where LLM API relays drop customer identity, causing prompt caches to no longer be isolated per downstream customer.

    Paper analysisRead analysis
  5. AttacksarXiv 2608.07430

    研究者提出 SN-Guided Diffusion,利用扩散大模型安全神经元实现离线越狱

    提出 SN-Guided Diffusion,利用安全神经元实现离线越狱

    Published
    Evaluated models
    Qwen2.5-7B-Instruct, LLaDA-8B-Instruct, Dream-Instruct-7B and others15
    Abstract论文揭示了扩散语言模型对自回归安全神经元的继承性,并提出纯离线黑盒越狱框架 SN-Guided Diffusion。

    论文系统评估了扩散大语言模型(DLLM)在安全对齐中的双重角色——既作为白盒攻击目标,又作为黑盒越狱生成的对抗工具。

    Paper analysisRead analysis
  6. AttacksarXiv 2607.23496

    Concept2Scenario:用 SAE 概念归因发现可迁移的越狱场景

    提出 Concept2Scenario,借助内部概念归因发现可迁移的越狱场景

    Published
    Evaluated models
    Qwen3.5-9B, Llama-3.1-8B, Ministral-3-8B and others6

    Abstract该研究从表征空间因果归因揭示场景削弱拒绝的机制,发现可复用脆弱场景先验,提升多模型黑盒越狱效能与探索速度。

    Paper analysisRead analysis
  7. AttacksarXiv 2605.07399

    研究者提出 GPO-V,用全局概率优化越狱扩散视觉语言模型

    提出 GPO-V,用全局概率优化的视觉扰动越狱扩散视觉语言模型

    Published
    Evaluated models
    LLaDA-V, LaViDA
    Abstract作者用视觉扰动改 dVLM 的全局生成概率,以绕过两类拒绝。

    GPO-V 把越狱从“开头必须顺从”改成“整段去噪概率朝肯定方向走”。。

    Paper analysisRead analysis
  8. AttacksarXiv 2604.24983

    PEO:直接优化原始提示词嵌入的 LLM 白盒越狱方法

    提出 PEO,直接优化原提示词嵌入实现白盒越狱

    Published
    Attacker
    White-box
    Evaluated models
    Vicuna-7B-v1.3, Llama-2-7B-Chat, Qwen3-4B-Instruct and others4
    AbstractPEO 原位优化提示词嵌入、多轮只重攻失败样本。

    PEO是一种不追加对抗 token 的 white-box 越狱:直接改已有提示词嵌入,并只对未成功样本更换续写目标再优化。

    Paper analysisRead analysis
  9. AttacksarXiv 2603.11132

    WebWeaver: Breaking the topology confidentiality of LLM multi-agent systems via covert context inference

    Propose WebWeaver, which compromises only a single agent and recovers the multi-agent communication topology from conversation context.

    Published
    Evaluated models
    Llama 3.1 8B, Qwen 2.5 7B, Mistral NeMo 12B and others6
    AbstractWebWeaver infers topology from context.

    WebWeaver is an inference attack against the confidentiality of LLM-MAS communication topology: it controls only one arbitrary agent and recovers the full graph from dialogue context, rather than querying agent IDs.

    Paper analysisRead analysis
  10. AttacksarXiv 2602.20595

    OptiLeak: Efficiently Reconstructing Prompts with Reinforcement Learning in Multi-Tenant LLM Services

    Proposing OPTILEAK, reconstructing others' prompts token-by-token via a KV cache side channel

    Published
    Attacker
    Black-box
    Evaluated models
    Qwen-2.5-3B-Instruct, Qwen-2.5-7B-Instruct, Qwen-2.5-14B-Instruct and others4

    AbstractOPTILEAK uses SFT and hard-token DPO to reduce the number of requests per token; the authors claim the risk of cache prompt leakage is higher than previously reported.

    Paper analysisRead analysis