Skip to content

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

Oct 12 · Monday

All papers

Found 6 papers
Filter4

Attack type

Jailbreaks6Prompt injection7Poisoning and backdoors3Malicious use0

Model risks

Deception and scheming4Misalignment and value divergence3Inappropriate refusal9Hallucination and fabrication3

Impact

Harmful content4

System type

Models and APIs5

Attack surface / entry point

User input6

Layer

Model layer4
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 255papers have no tags yet and are not included in filtered results.
Attack type：JailbreaksAttack type：Malicious useModel risks：Misalignment and value divergenceModel risks：Hallucination and fabricationClear all

Another 255papers do not have category tags yet and are not included in filtered results.

  1. DefensesarXiv 2610.07276

    SAFESHIELD：小语言模型部署时安全的决策组织框架

    提出 SAFESHIELD，组织小语言模型部署时的安全决策

    arXiv
    2610.07276
    Published
    Layer
    Application layer
    Setting
    LLM applications
    Evaluated models
    Qwen3-4B-Instruct
    AbstractSAFESHIELD 用可审计决策轨迹组织 SLM 部署时的四项安全决策，并报告协调依赖影响端到端结果。

    SAFESHIELD 是面向小语言模型的部署时安全框架，把运行时安全从单项审核改写成可检查的决策过程。

    Paper analysisRead analysis
  2. DefensesarXiv 2609.36657

    Constitutional adapters：用推理期干预防御越狱与失准

    提出 Constitutional adapters，蒸馏宪法行为以缓解越狱与失准

    arXiv
    2609.36657
    Published
    Layer
    Model layer
    Setting
    Models and APIs
    Evaluated models
    Qwen3.5-4B, Llama-3.1-8B-Instruct, Gemma-4-E4B and others4

    Abstract差分宪法适配器在长上下文和多轮上更稳，并可按 α 换取遵从。

    Paper analysisRead analysis
  3. Evaluation and benchmarksarXiv 2609.29429

    Jev 被用于零样本检测十类对齐失败，中位 AUROC 达 0.886

    提出 RLCDAlignBench，零样本检测多种对齐失败

    arXiv
    2609.29429
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Evaluated models
    Jev (jev-1.13.0)
    Abstract一次通用问题已能给多数对齐失败排序，大的误差来自状态和标签。

    RLCDAlignBench 测量用强化学习做校准决策的 Jev，能否当作对齐失败的零样本检测器。

    Paper analysisRead analysis
  4. Risky behaviorarXiv 2609.08186

    论文揭示大推理模型的对齐崩溃并提出 Reasoning Trap 越狱范式

    揭示推理加深引发的对齐坍塌，用 Reasoning Trap 放大越狱

    arXiv
    2609.08186
    Published
    Layer
    Model layer
    Setting
    Models and APIs
    Evaluated models
    Qwen3-8B, Qwen3-14B, Llama2-7B and others8
    Abstract更深推理提升题目准确率，同时提高扰动下的相对损失。

    作者研究扩展推理是否以对齐稳健性为代价，提出 ALR、RT 与 RRA。。

    Paper analysisRead analysis
  5. Evaluation and benchmarksarXiv 2608.20554

    aiXamine：跨维度评测 LLM 安全、安全（security）与隐私权衡

    提出 aiXamine 黑盒基准，统一评测 LLM 安全、稳健与隐私权衡

    arXiv
    2608.20554
    Published
    Layer
    Model layer
    Setting
    Models and APIs
    Evaluated models
    openai/gpt-5-mini, openai/gpt-4o, anthropic/claude-haiku-4.5 and others15
    AbstractaiXamine 用九服务黑盒协议比较百余个模型，作者称单轴信任度进展可能削弱其他轴。

    aiXamine 是一套黑盒评测协议，用来在同一条件下比较 LLM 的 safety、security 与 privacy，而不是再产出一个能力分。

    Paper analysisRead analysis
  6. Evaluation and benchmarksarXiv 2602.19450

    对面向可信执行环境的 Claude Opus 与 ChatGPT 安全顾问进行红队测试

    用 TEE-REDBENCH 评测 TEE 顾问的机制幻觉与提示诱导失败

    arXiv
    2602.19450
    Published
    Layer
    Model layer
    Setting
    Models and APIs
    Evaluated models
    ChatGPT-5.2, Claude Opus-4.6
    Abstract作者用 TEE-REDBENCH 测两款顾问，称失败可迁移且流水线控制可降低发生率。

    TEE-REDBENCH 红队评测被当作 Intel SGX 与 Arm TrustZone 安全顾问的 ChatGPT-5.2 和 Claude Opus-4.6，看架构推理、远程证明语义、缓解建议和策略约束下的拒答，并测提示诱导失败能否跨助手迁移。作者把助手看成随机策略。

    Paper analysisRead analysis
You have reached the end