1
韩国国产自杀式无人机海上试飞失败：两次起飞均数秒后坠海
IT之家 8 月 26 日消息，据新华社今日报道，韩国海军于 8 月 25 日在韩国南部海域进行了国产自杀式无人机（游荡弹药）的海上作战试验，两次放飞均以无人机坠海告终，试验未能完成预定目标。 此次测试在韩国南部海域进行，以大型运输舰“马罗岛”号（ROKS Marado，满载约 1.45 万吨）为平…
2
马斯克首次向 Cursor 全员讲话曝光：称 Grok 已经落后，而他“不习惯失败”
马斯克罕见内部讲话曝光：亲承Grok落后、披露600亿收购Cursor内幕，还预告用员工数据训练AI，野心与争议齐飞。
IT之家 8 月 25 日消息，当地时间 8 月 14 日，SpaceX 宣布完成了对 AI 编程初创公司 Cursor 的 600 亿美元 （IT之家注：现汇率约合 4,043.23 亿元人民币） 收购。 据 The Information 报道，就在官宣当天，埃隆 · 马斯克（Elon Musk…
3
Beyond End-to-End Success: Diagnosing Failures in Long-Horizon Security LLM Agents
系统化诊断长时程安全LLM代理的失败根因，突破端到端指标的盲区。
arXiv:2608.20563v1 Announce Type: cross Abstract: Long-horizon security LLM agents must carry information and decisions across many dependent interact…
4
Our AI reviewer invented a request. Our producer retried 245 times.
AI幻觉率仅0.2%，却让生产流程空转245次，一场关于AI可靠性的真实事故复盘。
We run ~100 LLM agents unattended on local models. Last week we found one document that had been rewritten 245 times in 5 days — every attempt rejecte…
5
Calibrating Criterion Revision in LLM Agents: Failure Modes and a Trace-Anchored Protocol
面对LLM代理频繁修订判断标准，这项研究揭示失败模式并提出追踪锚定协议，值得AI研究者细读。
arXiv:2608.20729v1 Announce Type: new Abstract: Language-model agents can improve after failure or carry text across episodes without revising what co…
6
杜撰某互联网企业“上市失败”，自媒体博主拼凑信息博眼球被行政拘留
IT之家 8 月 23 日消息，据法治天地今日报道，今年 7 月 8 日，某互联网企业向上海黄浦警方报案，称某自媒体账号发布内容为“知名企业上市失败”的不实贴文，对公司声誉造成负面影响。 黄浦公安分局网安部门接报后迅速开展调查，锁定了该账号运营者莫某。经查，莫某在某证券投资平台运营一个投资类账号，他…
7
Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution
不靠大模型推理，轻量图神经网络也能精准归因智能体失败，效率与可解释性兼得。
arXiv:2608.18575v1 Announce Type: new Abstract: Large language model (LLM)-based multi-agent systems (MAS) often exhibit complex failure modes, which …
8
Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution
给自主智能体戴上“安全紧箍”，行动边界+可信溯源+失败即停，杜绝越权操作。
arXiv:2608.16891v1 Announce Type: new Abstract: Agentic AI systems request tool actions that can modify files, send messages, launch jobs, or change w…
9
蓝箭航天朱雀三号总指挥戴政披露遥一回收失败原因：着陆点火出现异常燃烧，坠落在着陆场坪边缘
IT之家 8 月 19 日消息，蓝箭航天朱雀三号遥二运载火箭 今日在东风商业航天创新试验区发射升空 ，火箭一子级按预定程序成功着陆于甘肃省民勤县朱雀三号着陆场坪，飞行任务取得圆满成功。这是我国在重复使用火箭关键技术上取得的又一重大突破，朱雀三号成为我国首款成功入轨并实现陆地回收的运载火箭。 据央视新…
10
LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning Tasks
大模型能预判推理失败，却难选协作协议？这篇论文提出成本感知路由，高效平衡推理性能与开销。
arXiv:2608.14927v1 Announce Type: new Abstract: Multi-agent large language model (LLM) systems can improve reasoning by spending more computation, but…
11
Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents
别迷信Agent技能！实证研究发现，某些技能反而会拖累大模型任务成功率，反直觉结论值得关注。
arXiv:2608.11888v1 Announce Type: new Abstract: Agent skills are the de facto mechanism for extending LLM agents with reusable guidance. A skill can s…
12
LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations
不用等生成完，从预生成激活就能预判模型成败，给LLM装上“事前质检仪”。
arXiv:2602.09924v4 Announce Type: replace-cross Abstract: Running LLMs with extended reasoning on every problem is expensive, but determining which in…
13
LLM Reasoning for Subjective Tasks: Failure Modes, Mitigation, and Dynamic Reasoning Routing
LLM推理并非万能，主观任务中会失灵？这篇论文剖析失败模式并给出动态路由方案。
arXiv:2608.08889v1 Announce Type: new Abstract: Recommendation systems thrive on personalization, where ''correctness'' is rarely a binary truth but a…
14
Show HN: Whetstone – 20 Claude Code skills, each distilled from one real failure
从真实翻车中提炼的20条Claude Code技能，专治“看起来会了”的AI编程幻觉。
Article URL: https://whetstone.akbarsha.dev/ Comments URL: https://news.ycombinator.com/item?id=49235255 Points: 3 # Comments: 0
15
SearchAuditor: Auditing and Attributing Failures in Long-Horizon Search Agents
长时程搜索智能体频繁翻车？这个审计框架能精准定位并归因每一步失败。
arXiv:2608.05212v1 Announce Type: new Abstract: Deep search agents tackle challenging questions through long-horizon web interactions, a process that …
16
An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures
多语言环境下多智能体规划为何频频“翻车”？这篇论文给出可操作的失效诊断，直击任务关键信息丢失的根源。
arXiv:2608.03735v1 Announce Type: cross Abstract: Multilingual multi-agent systems exhibit substantial degradation beyond English, yet prior work rare…
17
The Tell-Tale Trace: Detecting Reasoning Failures in LLMs Using Chain-of-Thought Dynamics
从思维链的动态变化中识别大模型推理失误，为AI安全与可解释性提供全新监测路径，值得关注。
arXiv:2608.03291v1 Announce Type: new Abstract: Chain-of-thought (CoT) reasoning improves large language model (LLM) performance while also providing …
18
Training-Free versus Training-Based Intent Classification in LLMs: Accuracy, Robustness, and Failure Modes
大模型意图分类新研究：对比零样本与微调路线，揭示精度、鲁棒性与失败边界，值得关注。
arXiv:2608.02415v1 Announce Type: new Abstract: Intent classification in Large Language Models (LLMs) involves categorizing user prompts into predefin…
19
HPFA: Hypergraph-Based Paired Failure Attribution for LLM Reasoning
用超图建模LLM推理失败模式，成对归因精准定位错误根源，提升可解释性。
arXiv:2608.02026v1 Announce Type: new Abstract: Reflection is a powerful mechanism for LLM reasoning, yet its effectiveness hinges on accurately attri…
20
AdaMAST: Adaptive Failure Taxonomies for Improving LLM Agents
为LLM智能体构建自适应失败分类体系，系统提升多智能体场景的可靠性与调试效率。
Article URL: https://multi-agent-systems-failure-taxonomy.github.io/AdaMAST/blogs/adamast_paper/ Comments URL: https://news.ycombinator.com/item?id=49…