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

Oct 12 · Monday

All papers

Found 15 papers

Filter3

Model risks

Deception and scheming10 Reward hacking13 Misalignment and value divergence3 Dangerous agent actions19 Dangerous capabilities9 Inappropriate refusal2 Hallucination and fabrication4

Impact

Oversight failure12

System type

Models and APIs98 LLM applications3 AI Agent40 coding agent15 science agent7 Multi-agent systems5 Embodied systems and robotics6

Layer

Application layer12

System component

Monitors5

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 261papers have no tags yet and are not included in filtered results.

Model risks：Reward hacking Model risks：Inappropriate refusal System type：coding agent Clear all

Another 261papers 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
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
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
Risky behaviorarXiv 2610.03458Oct 2
论文：低监控读数不能证明行为受控
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
DefensesarXiv 2610.03055Oct 2
HackTrace：用生成状态监督检测代码生成中的奖励作弊
提出 HACKTRACE，用生成状态检测并抑制代码智能体奖励作弊
arXiv
2610.03055
Published
Oct 2
Layer
Application layer
Setting
coding agent
Evaluated models
Qwen3-8B, Llama-3.1-8B-Instruct, Qwen3.5-4B
- DefensesMonitoring and auditing
- RiskReward hacking
- ComponentMonitors
Abstract论文提出利用生成状态监督作弊尝试的 HACKTRACE 框架，实现高精度低延迟检测并在 GRPO 训练中有效抑制奖励投机。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.39533Sep 30
CATCH：面向编码 RL 奖励作弊的可控分析测试台
构建编码 RL 奖励作弊测试台 CATCH 并检验 CoT 监控失效
arXiv
2609.39533
Published
Sep 30
Layer
Model layer
Setting
coding agent
Evaluated models
Qwen3-4B, Qwen3.5-27B
- RiskReward hacking
- ComponentMonitors
AbstractCATCH 用双运行金标签研究编码 RL 的奖励黑客，并观察到监控会被注释适应削弱。
CATCH 是面向编码 RL 奖励黑客的可控分析测试床，用可利用环境、可调初始倾向和基于执行的金标签，在训练全程比较黑客动态与干预。
Paper analysisRead analysis
DefensesarXiv 2609.36170Sep 29
Assay：用内容寻址证据图约束 AI 编码智能体的声明失效
提出 Assay，用依赖锥哈希绑定代码断言并在合并门禁拦截失效声明
arXiv
2609.36170
Published
Sep 29
Layer
Application layer
Setting
coding agent
- DefensesPermissions and approval
- RiskDeception and scheming
- ComponentPermissions and sandboxing
AbstractAssay 提供随依赖内容失效的证据管理。
Assay 是一套面向 AI 辅助软件交付的内容寻址证据系统，关注代码改变后，测试通过、无秘密或行为保持等断言还能否用于批准合并。
Paper analysisRead analysis
Risky behaviorarXiv 2609.30266Sep 25
研究显示 Claude Code、Codex 等本地智能体可轻易篡改自身执行轨迹
测试编程智能体被诱导或自主篡改自身执行轨迹的行为
arXiv
2609.30266
Published
Sep 25
Layer
Application layer
Setting
coding agent
Evaluated models
GPT-5.6-Sol, GPT-6-Sol, Opus-5 and others11
- AttacksMalicious use
- RiskDeception and scheming
- ComponentMonitors
Abstract论文评估了本地智能体的轨迹防篡改能力，发现其可因外部指令或奖励追求而清除执行证据，亟需建立独立拦截记录机制。
这是一项针对本地大模型智能体执行轨迹完整性与安全防篡改边界的实证研究。
Paper analysisRead analysis
DefensesarXiv 2609.19101Sep 17
用内部表征监控与发现 LLM 评测中的奖励作弊
用内部表征探针监控并发现评测中的奖励作弊
arXiv
2609.19101
Published
Sep 17
Layer
Application layer
Setting
coding agent
Evaluated models
Kimi K3, GLM 5.2, Qwen 3.8 Max and others6
- DefensesMonitoring and auditing
- RiskReward hacking
- ComponentMonitors
Abstract论文分析了前沿开源模型在评测中的奖励投机，展示了利用内部激活均值差向量低成本监控与发现投机行为的可行性。
Paper analysisRead analysis
OtherarXiv 2609.12039Sep 11
现实才是最终验证者：智能体软件工程的两大关键缺口
提出双缺口框架，统一智能体软件工程中的奖励投机与幻觉
arXiv
2609.12039
Published
Sep 11
Layer
Application layer
Setting
coding agent
- RiskReward hacking
Abstract双缺口解释内环为何会放行部署后不可接受的实现，外环用现实证据持续修订 R、M 与 E。
Reality Is the Final Verifier 用 two-gap framework 解释智能体软件工程的主要失败，并主张用外层保证环持续收窄缺口，而不是试图在部署前证明缺口已关闭。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.08149Sep 8
研究者发布 SWE-Bench Pro Verified，修正奖励作弊与任务质量问题
发布 SWE-Bench Pro Verified，隔离答案通道并修订任务以抑制奖励破解
arXiv
2609.08149
Published
Sep 8
Layer
Application layer
Setting
coding agent
Evaluated models
GPT-5.6-Sol, Kimi-K3, GLM-5.3 and others7
- RiskReward hacking
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2608.22103Aug 23
Hack-Verifiable Terminal Bench：评测终端任务中的奖励作弊
提出 HVTB，在终端编程任务植入蜜罐以测量奖励作弊
arXiv
2608.22103
Published
Aug 23
Layer
Application layer
Setting
coding agent
Evaluated models
claude-opus-5, gpt-5.6-sol, gemini-3.1-pro and others5
- RiskReward hacking
Abstract用蜜罐测量终端任务上的 reward hacking，并比较五档提示词。
HVTB 用来在真实终端与编程任务上测量 reward hacking：智能体读取被植入的参考解答或验收测试，从而偏离任务意图。作者认为这一失败模式会随智能体更自主而更重要，且事后人工或 LLM judge 可能失效。
Paper analysisRead analysis
InterpretabilityarXiv 2606.05396Jun 4
研究用消融移除代码 LLM 拒答方向，分离拒答与生成能力
用 abliteration 消融代码 LLM 拒答方向，分离拒答与漏洞注入能力
arXiv
2606.05396
Published
Jun 4
Layer
Model layer
Setting
coding agent
Evaluated models
Qwen2.5-Coder-Instruct 3B, Qwen2.5-Coder-Instruct 7B, Qwen2.5-Coder-Instruct 14B
- RiskInappropriate refusal
Abstract拒答可用低秩编辑去掉，CWE-89 注入是否成功仍由模型规模决定。
这项初步可行性研究用 abliteration 把代码 LLM 的拒答和 CWE 注入能力拆开：意愿可以被低秩编辑解锁，注入是否成功仍看参数规模。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2605.19722May 19
研究用 30 项漏洞分析任务评测自主安全智能体的安全对齐效果
评测自主安全智能体在授权漏洞分析中的对齐拒答效应
arXiv
2605.19722
Published
May 19
Layer
Application layer
Setting
coding agent
Evaluated models
Gemma 4 31B official instruction-tuned model, TrevorJS Gemma 4 31B uncensored GGUF, Gemma 4 26B A4B official instruction-tuned model and others8
- RiskInappropriate refusal
Abstract轨迹级评测把拒答、接地、工具接口分开。
这项工作把安全对齐效应测成自主安全智能体的轨迹属性，而不是单轮拒答率。
Paper analysisRead analysis

You have reached the end

AI Safety Daily

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