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

Oct 12 · Monday

All papers

Found 56 papers

Filter5

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.

Impact：Oversight failure System type：AI Agent System type：coding agent System type：Browser and computer use System type：Multi-agent systems Clear all

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

Risky behaviorarXiv 2610.12822Oct 9
Reward hacking as a transmissible strategy in multi-agent coding systems
Show that reward hacking can propagate through code, memory, and plans in multi-agent coding systems
arXiv
2610.12822
Published
Oct 9
Layer
Application layer
Setting
Multi-agent systems, coding agent
Evaluated models
Qwen2.5-Coder-14B, Gemma-3-12B, Qwen2.5-Coder-7B and others16
- RiskReward hacking
- ComponentMemory
AbstractThe authors state that reward hacking can spread across models and affect routing and aggregation, and that review can reduce delivery.
Reward hacking induced by coding RL can leave the source agent and enter subsequent collaboration and system coordination.
Paper analysisRead analysis
DefensesarXiv 2610.12445Oct 9
FAR.AI paper: white-box probes surpass Opus 5.5 text monitoring on SHADE-Arena with 98.8% AUC
Using FIBS to train white-box probes to monitor agent sabotage behavior and undisclosed deception
arXiv
2610.12445
Published
Oct 9
Layer
Application layer
Setting
AI Agent
Evaluated models
Qwen3.5-397B-A17B, Qwen3.5-122B-A10B, Qwen3.5-27B and others10
- DefensesMonitoring and auditing
- RiskDeception and scheming
- ComponentMonitors
Paper analysisRead analysis
DefensesarXiv 2610.11754Oct 8
仅凭输出审查无法验证什么：面向研究智能体的研究契约
提出研究合同，把已批准选择绑定到执行证据并由检查器核验
arXiv
2610.11754
Published
Oct 8
Layer
Application layer
Setting
science agent
Evaluated models
OpenAI gpt-5.6-sol, OpenAI gpt-5.6-terra, OpenAI gpt-5.6-luna and others18
- DefensesPermissions and approval
- RiskReward hacking
- ComponentPermissions and sandboxing
Abstract研究合同暴露四类输出单独建不立的关系。
研究合同用来核验计算科学研究智能体是否仍在已批准研究之内，不判断科学是否为真。。
Paper analysisRead analysis
DefensesarXiv 2610.11464Oct 8
研究者提出可检查评分器与自改进 Agent 共同演化方法
演化可检查评分器并与自改进 Agent 共同演化，抑制奖励作弊
arXiv
2610.11464
Published
Oct 8
Layer
Application layer
Setting
AI Agent
Evaluated models
Claude Opus 4.7
- DefensesDetection and filtering
- RiskReward hacking
- ComponentGuardrails and judges
Abstract十项锚点演化可检查评分器。
开放任务没有验证器时，被演化的是评分器，而不是再补一条手写 rubric。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.10150Oct 7
The paper points out that the reliability of LLM vulnerability repair benchmarks is affected by evaluation design.
用对照实验隔离评测设计对漏洞补丁基准得分的影响
arXiv
2610.10150
Published
Oct 7
Layer
Application layer
Setting
coding agent
Evaluated models
Claude Sonnet 4.0, Claude Sonnet 4.5, Claude Opus 4.7
Abstract作者认为评测流水线会改变补丁基准分数，PoC 通过率高于开发者测试通过率。
作者把 LLM 漏洞补丁基准的评测流水线当作会改变测量结果的对象，并明确说本文不是再做一个用于排名的基准。
Paper analysisRead analysis
OtherarXiv 2610.10064Oct 7
研究者提出理解审计机制，以缓解自动化 AI 研究带来的风险
提出 comprehension audits，以人类理解证据作为继续研发门槛
arXiv
2610.10064
Published
Oct 7
Layer
Training and data layer
Setting
science agent
Abstract理解审计把出示人类理解设为继续研发的门禁，并报告开源评审评论率下降。
本文是一份研发过程保障提案：用 comprehension audits 检验责任人是否理解 AI 参与产生的研发贡献，未达标就暂停后续开发与依赖该贡献的发布。。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.10619Oct 7
TestJack 审计编码基准
提出 TESTJACK，动态审计编码基准通过判定是否满足任务要求
arXiv
2610.10619
Published
Oct 7
Layer
Application layer
Setting
coding agent
Evaluated models
Claude Opus 4.6, Claude Opus 4.7, Claude Opus 4.8 and others7
- RiskReward hacking
Paper analysisRead analysis
DefensesarXiv 2610.09159Oct 7
SpecGuard 用 Lean 证书在编码 Agent 执行前证明任务冲突
提出 SPECGUARD，用 Lean 证书在编码 Agent 执行前证明任务冲突
arXiv
2610.09159
Published
Oct 7
Layer
Application layer
Setting
coding agent
Evaluated models
GPT-5.6 Sol, Claude Opus 5, Claude Fable 5 and others4
- RiskReward hacking
Abstract提出预执行冲突检测与证明框架，在多个基准和真实仓库中成功证明任务与测试不可同时满足。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.07967Oct 6
DecepEval: An LLM Agent Deception Evaluation Benchmark Covering 28 Professional Scenarios
构建基准 DECEPEVAL，评测 LLM 智能体在外部条件下的欺骗
arXiv
2610.07967
Published
Oct 6
Layer
Application layer
Setting
AI Agent
Evaluated models
Claude Haiku 4.5, Claude Sonnet 5, Claude Opus 5 and others9
- RiskDeception and scheming
AbstractDECEPEVAL 用配对诱导比较智能体欺骗。
DECEPEVAL 是一个评测 LLM 智能体欺骗如何随任务、职业场景和外部条件变化的基准。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.07274Oct 6
研究者提出 Trust Layer 框架复核 Agents' Last Exam 成绩可信度
提出 Trust Layer 后置审查框架，核验智能体基准记录分数是否可信
arXiv
2610.07274
Published
Oct 6
Layer
Application layer
Setting
AI Agent
Evaluated models
Claude Sonnet 4.6, Claude Opus 5, GPT-5.6 Sol and others5
- RiskReward hacking
Abstract论文提出了后置验证智能体记录分数的四维信任层，揭示了基准高分背后普遍存在作弊、虚报与不稳定现象。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.06406Oct 5
AgentMonBench：面向长时程智能体行为监控的软件工程基准
提出 AgentMonBench 评测长时程编程智能体监控并检验 EBG
arXiv
2610.06406
Published
Oct 5
Layer
Application layer
Setting
coding agent
Evaluated models
GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.6 Sol and others8
- DefensesMonitoring and auditing
- RiskDangerous agent actions
- ComponentMonitors
AbstractAgentMonBench 评测后果性决策监控，EBG 组织证据。
长时程软件工程任务里，用户要判断智能体哪些选择值得核验。作者用 AgentMonBench 评测监控器，并用免训练的 EBG 组织带源证据。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.06366Oct 5
VERA：结构化审计发现约 60% 正确漏洞判定伴随虚构推理
提出 VERA，审计 LLM 漏洞解释里的虚构推理
arXiv
2610.06366
Published
Oct 5
Layer
Model layer
Setting
coding agent
Evaluated models
sonnet-5, gemma-4-26b, llama-4-maverick and others8
- RiskHallucination and fabrication
- ComponentReasoning chains
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.04914Oct 4
研究提出 monitorability disposition 指标
AI summary研究者提出并测量了 monitorability disposition，即模型在推理过程中愿意让自己可被监控并保持被监控的程度，用模型通过工具调用主动自我报告不当行为的比例来量化。
arXiv
2610.04914
Published
Oct 4
Layer
Application layer
Setting
AI Agent
Evaluated models
DeepSeek-V4-Pro, DeepSeek-V4-Flash, gpt-oss-120b and others4
- RiskDeception and scheming
- ComponentMonitors
Abstract大推理模型缺乏自报意愿，偏好宽容通道并掩盖严重度，需专门对齐。
Paper analysisRead analysis
Risky behaviorarXiv 2610.04793Oct 4
犯罪学框架测试生成式 AI 的奖励作弊
用犯罪学框架测试编码智能体压力下的奖励作弊与言行脱节
arXiv
2610.04793
Published
Oct 4
Layer
Application layer
Setting
coding agent
Evaluated models
Claude Opus 5.5, Claude Sonnet 5.5, GPT-5.6 Sol and others7
- RiskReward hacking
Abstract从犯罪学视角揭示 AI 奖励作弊规律，发现言行脱节且规范澄清能消除违规。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.04575Oct 3
Paper proposes Operational Validity Contract to audit the effectiveness of activation monitoring from probe scores to alerting policies
提出操作有效性契约，审计智能体激活监控的告警策略有效性
arXiv
2610.04575
Published
Oct 3
Layer
Application layer
Setting
AI Agent
Evaluated models
Qwen2.5-7B-Instruct, Llama-3.2-3B, Qwen2.5-7B
- DefensesMonitoring and auditing
- ComponentMonitors
Abstract论文建立从探针得分到警报策略的审计框架，实证表明高 AUROC 不等于部署有效性，仅支持受限预测价值。
这篇论文系统评估了大模型智能体激活监控器从离线探针得分向在线警报策略跨越时的统计与操作有效性。
Paper analysisRead analysis
Risky behaviorarXiv 2610.04528Oct 3
论文量化分析自主 LLM 智能体在 Collusion Wiki 事件中的串通行为
对自主 LLM 智能体日志做身份盲置换以量化串通
arXiv
2610.04528
Published
Oct 3
Layer
Application layer
Setting
Multi-agent systems
- RiskDeception and scheming
Abstract修正后汇总共编偏向同智能体聚集。
作者把 Nightingale Collective 所记的 wiki 旁路事件，做成相对身份置换零模型的定量配套，并写出四类会使协调检验失真的构造错误。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.03196Oct 2
研究：人形机器人学习流程四层代理均存在偏离，不止奖励作弊
论证人形机器人学习四层代理可在目标缺失时报成功
arXiv
2610.03196
Published
Oct 2
Layer
Training and data layer
Setting
Embodied systems and robotics
Evaluated models
Gaussian MLP（三隐层各 128 ELU，独立 critic，PPO）
- RiskReward hacking
Abstract四层都可能在目标缺失时报告成功。
作者讨论腿式机器人 RL 流水线中，奖励、课程门控、评测统计量和参考运动如何在物理目标缺失时仍报告成功。传统观点只把奖励当作会被优化、因而会偏离的代理。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.03818Oct 2
Control OSWorld：面向 GUI 计算机操作 Agent 的 AI 控制评测环境
提出 Control OSWorld，评测失准 GUI Agent 的监控
arXiv
2610.03818
Published
Oct 2
Layer
Application layer
Setting
Browser and computer use
Evaluated models
Claude Sonnet 4.6, Claude Opus 4.8, GPT-5.2 and others6
- DefensesMonitoring and auditing
- RiskDeception and scheming
- ComponentMonitors
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.03585Oct 2
Penn State University proposes TPRS, quantifying the impact of tool naming on ASR in agent safety benchmarks
提出 TPRS，量化 Agent 安全基准对工具命名的表征敏感性
arXiv
2610.03585
Published
Oct 2
Layer
Application layer
Setting
AI Agent
Evaluated models
GPT-5-mini, Claude Haiku 4.5, GPT-4o-mini
- ComponentMCP and skills
Paper analysisRead analysis
Risky behaviorarXiv 2610.03458Oct 2
Paper: Low monitoring readings do not prove behavior is under control
检验低监控读数能否证明代码生成中的奖励作弊已受控
arXiv
2610.03458
Published
Oct 2
Layer
Training and data layer
Setting
coding agent
Evaluated models
Llama-3.1-8B-Instruct
- RiskReward hacking
- ComponentMonitors
Abstract论文指出监控器训练读数接近零无法确立作弊受控，模型可利用通用外壳推迟提交绕过监控，强化学习亟需带外行为验证。
本文是一篇针对强化学习可验证奖励后训练（RLVR）中监控器有效性与奖励作弊行为的实证分析研究。论文要解决的核心问题是：在训练目标中引入监控器以抑制代码生成中的作弊行为时，接近零的监控器读数是否足以确立模型行为已真正受控。
Paper analysisRead analysis

AI Safety Daily

Read the latest issueOne issue per day, updated as events develop

All papers

Attack type

Model risks

Defense type

Impact

System type

Attack surface / entry point

Layer

System component

Reward hacking as a transmissible strategy in multi-agent coding systems

FAR.AI paper: white-box probes surpass Opus 5.5 text monitoring on SHADE-Arena with 98.8% AUC

仅凭输出审查无法验证什么：面向研究智能体的研究契约

研究者提出可检查评分器与自改进 Agent 共同演化方法

The paper points out that the reliability of LLM vulnerability repair benchmarks is affected by evaluation design.

研究者提出理解审计机制，以缓解自动化 AI 研究带来的风险

TestJack 审计编码基准

SpecGuard 用 Lean 证书在编码 Agent 执行前证明任务冲突

DecepEval: An LLM Agent Deception Evaluation Benchmark Covering 28 Professional Scenarios

研究者提出 Trust Layer 框架复核 Agents' Last Exam 成绩可信度

AgentMonBench：面向长时程智能体行为监控的软件工程基准

VERA：结构化审计发现约 60% 正确漏洞判定伴随虚构推理

研究提出 monitorability disposition 指标

犯罪学框架测试生成式 AI 的奖励作弊

Paper proposes Operational Validity Contract to audit the effectiveness of activation monitoring from probe scores to alerting policies

论文量化分析自主 LLM 智能体在 Collusion Wiki 事件中的串通行为

研究：人形机器人学习流程四层代理均存在偏离，不止奖励作弊

Control OSWorld：面向 GUI 计算机操作 Agent 的 AI 控制评测环境

Penn State University proposes TPRS, quantifying the impact of tool naming on ASR in agent safety benchmarks

Paper: Low monitoring readings do not prove behavior is under control