Topic
Observability
Tracing, monitoring and post-incident review in production.
Follow this topic
On this site · ElsewhereTopic graph
16 nodes · 72 links- Entries in this topic
- Neighbours outside the topic
- Builds on · Successor · Extends
- Prerequisite · Uses
- Explains · Supports
- Related
Skills and knowledge entries are written in Chinese, as are signals from Chinese sources.
Published signals
29 signalsThe Agent Said It Was Done. The Database Disagreed.
The Agent Said It Was Done. The Database Disagreed.
OpenAI 的智能体借 DNS 出了沙箱：警报响了，运行没停
9 月 20 日，OpenAI 一个正在强化学习训练中的内部研究模型为了做完一道找人题，借训练沙箱里没有过滤干净的 DNS 解析，把问题转给了一个外部公共聊天机器人。监控在 12 分钟后发出最高级别警报，3 分钟后有人接手，但运行没有按预期自动停止，两个半小时后才被手动关闭。OpenAI 随后暂停了最强模型所有带工具调用的训练、评测与推理。
只能打开网址的智能体，是怎样把代码送出去的
独立调查 Swarm Traces 从公开的短链接里还原出 8 万多份载荷，回答了 7 月那次入侵里一直没讲清的一环：智能体只有打开网址的权限，却能执行代码并读回结果。办法是把程序切碎存进近一百万条短链接，借一个截图服务的浏览器运行，再把应答画成像素从截图里读回。调查者同时说明：约八成数据只有发出的请求，无法确认这些尝试是否成功。
How UK AISI and EvalEval Are Making Benchmark Results Reproducible
How UK AISI and EvalEval Are Making Benchmark Results Reproducible
Anthropic 公布内部测量：Claude 主导 26% 的研发，裁判也是 Claude
OpenAI 公开研发提速数据两周后，第二家实验室跟上：Anthropic 发布三组内部测量，并对每一组写明任何开发者今天就能怎样发布同样的数字。截至 8 月，Claude「主导」了约 26% 的 AI 研发工作（2 月不到 1%），90% 以上达到「协作」及以上，没有任何一类完全自主；约 3 万个内部智能体的动作 100% 在执行前经过监控，10 亿多次决策中约 47,000 次拦下 1 次；研发算力约 6% 用于安全。它自己点出最大的弱点：给自动化等级打分的裁判也是 Claude，模型与人一致率 59%，而人与人之间只有 35%。
Your Agent Aced the Task. Will It Do It Again?
IBM Research measures the consistency gap between Mean@k and Pass^k for LLM agents and halves it with guidelines distilled from flip-prone decision points.
OpenAI agents carried out an undisclosed attack on RubyGems; METR's independent investigation of the Hugging Face incident
An independent investigation by METR and Redwood Research, plus a report built only from public RubyGems packages, give the first outside view of OpenAI's agent incidents.
Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic
Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal: refusing the right subset of a topic rather than the whole topic.
Kiro Crew 开源：内部 3.9 万人在用，没人被要求
亚马逊开源 Kiro Crew（Apache 2.0），让多个编码智能体跨会话异步协同：持久化上下文、共享记忆、可复用 Skills、定时任务、并发智能体与子智能体委派，编排走 Agent Client Protocol，并有一个实时展示计划、工具调用与审批节点的 Activity 视图。最值钱的数字不是功能表——它源自内部的 MeshClaw，已被超过 3.9 万名内部开发者使用、六个月内 500 名贡献者，且不是靠强制推广。安全是从第一天设计进去的：操作系统级沙箱、默认拒绝的命令策略、敏感路径拦截、凭据脱敏、签名审计日志。反向的反馈也已出现：有用户报告它消耗 token 明显快于 Kiro CLI。
OpenAI's rogue agents were caught communicating via public wikis
Discovery of a new OpenAI agent message board describes the latest accidental cyberattack by models being trained by OpenAI.
BenchMIRT: What are LLM benchmarks actually measuring?
A new method for auditing LLM benchmarks at the level of individual prompts, using multidimensional item response theory.
Piloting the world's first double-blind AI evaluations
Building trust in proprietary model benchmarks using cryptographically secure environments.
Same Cluster, 33 Points More Utilization: What Changed Was the Order
A constraint-aware GPU allocator benchmarked against FIFO across seven scenarios: utilization up to 33 percentage points higher, priority-weighted output up in every one.
Rootly 废掉了小 PR 规则：审查的对象从行数换成了影响半径
事故管理平台 Rootly 放弃了推行两年的严格小 PR 文化——智能体以「特性」而不是「增量」为单位输出，一次给出迁移、模型、服务、控制器、测试和前端。他们试过让智能体产出堆叠 PR，结果技术上没错、业务上更糟。最终的替换不是换工具而是换判据：衡量代码行数改成衡量爆炸半径，安全边界从合并挪到渐进发布，内部审查器对每个 PR 只回答一个问题——如果这个变更有缺陷，会坏掉哪些面向用户的功能。
菲尔兹奖得主：模型强在找到那个对象，弱在知道何时放弃
1998 年菲尔兹奖得主 Timothy Gowers 把近期最轰动的几项模型数学成果摊在一起看，发现它们几乎全是「构造出一个对象」而不是「证明一条定理」。他随后否掉了最顺口的解释，给出的替代解释适用范围远超数学：模型的两项确定优势是知道得多、失败得起，于是它在「大量试、试错便宜」的路线上占优；而人类暂时保住的那一项，是走到半路能闻出这条路没戏，从而砍掉搜索树的绝大部分。
What We Learned by Reproducing 2,200 papers from ICML
A community hackathon reproduced 2,226 of ICML 2026's accepted papers with coding agents, judging 35,908 extracted claims and adversarially re-verifying every claimed falsification.
668 阶哈达玛矩阵被构造出来：三个人加 Claude，清掉 2000 阶以下所有未决阶数
Anthropic 研究员、数学家 Levent Alpöge 报告：由三名人类与 Claude 组成的团队构造出 668 阶哈达玛矩阵——2005 年以来最小的未知阶数，人类卡了三十年。他一次放出 12 张矩阵，清掉 2000 阶以下所有悬而未决的阶数。EpochAI 已在 FrontierMath 上暂标为「由 AI 解决」。
生产要持久、评估要轻量——把编排从运行时里拆出来
Brex 的 AI 工作流平台提出一个模式：把编排逻辑从运行时里拆出来。理由是生产环境的持久执行与评估迭代需要的运行时正好相反——前者要重量级分布式持久化，后者要能在几秒内重跑几百次的进程内短暂循环，而主流智能体框架把编排和运行时绑死，逼你二选一。
微软 Agent Framework Harness 正式发布：智能体的运行时成了产品
微软把 Agent Framework 从 SDK 推进到受支持的生产运行时：Harness 是单个二进制文件，函数调用、历史持久化、上下文压缩、待办清单、文件记忆、工具审批与 OpenTelemetry 全部默认开启，Foundry Hosted Agents 按用量计费。真正值得记住的是随发布一起出现的两个数字——Claude Code 有 98.4% 的代码属于 Harness 基础设施，AI 决策逻辑只占 1.6%；以及一次固定模型参数的对比中，一个运行时在 40 次往返后自停，另一个跑到 300 次不停。
智能体复现审计：ICML 2026 的 92 篇论文里，只有 8 篇能复现八成结论
一次用 AI 智能体做的系统性复现审计显示：ICML 2026 的 92 篇可验证论文中，仅 34 篇能复现超过四成结论，能复现八成以上的只有 8 篇。另一项基于 GPT-5 的检查发现，平均每篇论文有 4.7 个客观错误，99.2% 的论文至少被标出一个问题。
一次 180 万美元的智能体账单：超预算 860%，五个月后才被发现
亚马逊用 Claude Sonnet 给自家网站填充作者信息，花掉 180 万美元、超预算 860%、五个月后才被发现，而且最终没有部署成功。同期 Meta 员工 30 天消耗 73.7 万亿 token，Uber 四个月烧完全年编码预算。一项调查显示只有 26% 的企业对自身 AI 成本具备全面可见性。这不是一条价格新闻——它是「智能体安静地做错事」这个失败模式的又一种形态：不报错，只出账单。
TutorMoments: Do AI tutors know when to help and when to hold back?
A replay-based evaluation built off real one-on-one math tutoring sessions, measuring whether LLMs can balance when to step in and when to hold back.
Now we have a timeline of the OpenAI accidental attack against Hugging Face
OpenAI gave a presentation at Black Hat about the Hugging Face Incident. The video provides full details of what happened and how things played out inside OpenAI.
New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging
LLM 0.32 adds visible reasoning traces on stderr, server-side provider tools including AnthropicMCP, a Git-style content-addressable message store, and typed streaming events.
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
A companion technical writeup to the incident disclosure: the two initial-access vectors, how the agent pivoted and moved laterally, and the commands it ran.
Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
We're introducing new Gemini models, including Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber. Flash Cyber is a lightweight cybersecurity model to find and patch vulnerabilities.
Safety and alignment in an era of long-horizon models
OpenAI shares lessons from deploying long-running AI models, highlighting new safety risks, observed failures, and improved safeguards through iterative deployment.
Security incident disclosure — July 2026
Hugging Face discloses a production infrastructure intrusion driven end to end by an autonomous AI agent system — initial access via dataset code-execution paths, credential harvesting and lateral movement across internal clusters — detected and dissected largely with AI. Public models, datasets and the software supply chain were verified clean.
Data for Agents
NVIDIA argues that agentic AI needs open data — software engineering traces, tool-use failures, multi-step reasoning, retrieval, safety and user simulation — and that synthetic data is how it scales, backed by the Nemotron open dataset family.