Skip to content

All news

51 today

Oct 22Thursday

The Decoder

An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks

Symantec安全研究人员发现,攻击者利用未受保护的语言模型驱动攻击智能体,由模型所有者承担算力费用。蜜罐截获的完整提示词包含五层越狱设计和74个攻击模块,“Skynet模式”授权连续执行最多100,000步,计划循环扫描、入侵和安装挖矿程序。蜜罐未执行这些工具,分析也未说明该越狱能否有效突破当前商业模型的安全防护。

MIT Technology Review · AI

Making the case for context-aware AI

Celonis 企业 AI 首席科学家 Devavrat Shah 提出“决策智能”,主张 AI 智能体必须理解任务在业务流程中的角色与约束,而非在真空中执行。他建议企业从具体业务决策切入,找出最难的环节测试 AI 能否提速并衡量价值,再决定是否规模化。缺乏运营上下文会让智能体忽视约束、犯错或消耗过多算力,落地还需配套员工教育。

Hacker News front page

Using LLM for confidence scoring is a bad idea

LLM 自报的置信度是逐个 token 生成的文本,文章指出,这些数字缺乏针对实际任务正确率的校准依据。可靠的置信度需要利用预测所依据的信息,并通过同类问题的实际结果持续检验,单纯符合历史平均正确率还不够。文章以客服智能体为例,提出记录检索命中、重复回答一致性和人工复核结果,按历史正确率建立评分规则并定期检查校准情况。

Hacker News front page

Achieving AI-style reasoning without an LLM

Minotaur 是一个不依赖语言模型和 GPU 的 AI 架构,基于 ELC 通过测量而非预测来推导答案,在单个普通 CPU 核心上每题约 0.25 秒完成演绎、归纳、溯因、诊断、信念修正等推理。在未见其代码的人设计的盲测中,它在提问会改变证人记忆的警察审讯场景,以及仅凭数据学习、诊断和预测系统等任务上,表现与当前用文字推理的语言模型相当。

Hacker News front page

Show HN: The fastest browser for AI agents to run browser automation

ego lite 为 AI 智能体提供独立的 Spaces 工作区,支持多个浏览器自动化任务并行运行,不干扰用户当前标签页。它通过 ego-browser 接入 Claude Code、Codex、Cursor 等智能体 CLI,让智能体直接编写 JavaScript 函数操作浏览器,并称复杂任务完成速度最高提升至 2.5×。

Latent Space

OpenAI and Wiz each find PHP2Shell flaw in PHP core, coordinating fix for about two months

OpenAI's Patch the Planet project helped find PHP2Shell, a vulnerability in the PHP core, and took part in a coordinated fix lasting about two months. Researchers at Calif used GPT-5.6 Sol; Wiz reported finding the same flaw independently. The researchers demonstrated unauthenticated remote code execution against WordPress, Drupal and Joomla.

Why it matters: The roughly two-month coordinated PHP2Shell fix splits work between maintainer patches and vendor stopgaps, which sets where responsibility sits in vulnerability response.

Hacker News front page

Claude plugs into Google Docs, Sheets and Slides, public beta open to all paid plans

Claude for Google Workspace is in public beta on every paid Claude plan. It sits in a sidebar in Google Docs, Sheets and Slides, reads the open file and edits it directly.

Why it matters: Claude can now change documents, spreadsheets and slides in place, with each edit approved item by item.

Hacker News front page

A Markdown File Should Just Open

作者指出,Markdown 诞生已超 20 年,却仍无法在大多数电脑和手机上直接以文档形式打开,用户收到 .md 文件往往只能看到原始语法或另装软件。Apple 已在 Notes 中加入 Markdown 导入导出,但导入会转成富文本格式,并非原地打开编辑;Windows 11 记事本已支持 Markdown 格式与源码切换。

AI HOT · Products

MiniMax Code desktop adds Mini App for long-running agent apps

MiniMax Code's desktop client now has Mini App, which moves agent work into a continuously running front-end and back-end app. Local sessions are supported on macOS and Windows.

Why it matters: Agent work moves into a persistent Mini App, so data updates, saved results and manual edits sit on one page instead of being hunted down across past conversations.

AI HOT · Products

Kling 4.0 已在 OnSolo 上线,支持单次生成 30 秒原生视频

可灵(Kling)宣布 Kling 4.0 已在 OnSolo 上线,支持单次生成 30 秒原生视频。模型支持在快速动作和大运镜下保持稳定,提供立体声、更准确的口型同步及 21:9 宽屏。单个任务可叠加最多 15 个参考图和 10 个关键帧,以保持角色、镜头与剧情一致。

Hacker News front page

WorldTUI: I trained a 360M model to pretend to be a terminal

WorldTUI 基于 Qwen2.5-Coder-0.5B 微调,用 360M 参数模型根据按键预测终端屏幕变化,背后没有真实 shell 或文件系统。训练使用 336k 段真实终端会话,并通过错误修复训练改善一致性,通过并行生成屏幕行和顺序验证加速推理。

Hacker News front page

OpenAI CLI

OpenAI CLI 是 OpenAI REST API 的官方命令行工具,支持响应创建、文件管理、图像生成及管理端接口调用。可通过 Homebrew 或 Go 安装,成功响应默认输出可读文本,脚本解析 API 数据时需显式使用 --format json 或 --format jsonl。

AI HOT · Industry

NSA 就 GPT-5.6 Sol 发现的 PHP2Shell 关键漏洞发布修复警告

NSA 就 Calif 研究人员使用 GPT-5.6 Sol 发现的两个 PHP 关键漏洞(PHP2Shell)发出警告,敦促组织打补丁或将未修补服务器下线。两个漏洞位于 PHP 核心,在 Apache mod_php 上演示的利用可在应用代码运行前触发。

Hacker News front page

I can't believe PowerShell is good

微软的 PowerShell 7(PowerShell Core)是基于 .NET Core 的跨平台版本,可在 Windows、Mac 和 Linux 上免费安装使用,Windows 11 仍只自带旧版 Windows PowerShell。

AI HOT · Industry

Anthropic may file S-1 publicly as soon as Friday, with IPO possibly larger than SpaceX

Yahoo Finance, citing three people familiar with the matter, says Anthropic may publicly file its S-1 as soon as Friday, moving toward a planned November listing. The company confidentially filed a draft IPO application in June. The public filing would disclose financial data and potential risks, though timing still depends on market conditions. Wall Street expects the IPO to raise up to $100 billion at a valuation of up to $2 trillion, possibly larger than SpaceX's June IPO.

Why it matters: A public S-1 would give ordinary investors Anthropic's financial data and risks, and the listing tests how much appetite the market has for AI.

AI HOT · Industry

PHP2Shell 事件复盘:AI 辅助发现 PHP 漏洞,补丁发布前泄露

PHP2Shell 事件涉及两个评分为 9.2 的 PHP 远程代码执行漏洞,在补丁发布前泄露,并引发多国政府网站预防性下线。漏洞编号为 CVE-2026-109618 和 CVE-2026-109619,攻击演示视频也已外流。OpenAI 表示,这些漏洞由 GPT-5.6 Sol 辅助发现。

AI HOT · Industry

Bloomberg reports DeepSeek's new funding round at about $15 billion

Bloomberg reports DeepSeek's latest round at about $15 billion, at a pre-money valuation of about $74.8 billion. The funding target was raised at least twice during talks, from about $7.5 billion to $12 billion. CATL and Tencent are among the largest investors, with CATL reportedly putting in about $1.5 billion. With a first round of more than $7.5 billion completed in June, DeepSeek has raised about $22.5 billion in roughly six months.

Why it matters: DeepSeek raised about $22.5 billion in roughly six months, and its new round's target rose from about $7.5 billion to $12 billion during talks.

AI HOT · Industry

WSJ:马斯克希望以 AI 为核心推动 SpaceX 与 Tesla 合并

WSJ 报道,马斯克希望以 AI 为核心,将估值约 $2T 的 SpaceX 与约 $1.5T 的 Tesla 合并。SpaceX 的 AI 业务已成其最大业务,Q3 营收为 $7.56B,一年前为 $813M,季度 $19.6B 资本开支中超过 $16.4B 投向 AI。材料称 AI 业务推动 SpaceX 实现季度利润 $4.4B,上年同期亏损 $1.05B。

Computing Life · Share · Yage

Why multi-agent AI costs more: what the extra budget buys

Multi-agent setups spend more tokens to gain specific engineering benefits, and whether a task gains depends on its parallelism, how much information it must cover, and its context state. Drawing on Anthropic, Cognition and several studies, the article sorts the gains into four kinds: shorter end-to-end time, gap-filling, more search diversity and context isolation. Team cost is usually several times that of a single agent.

Why it matters: It breaks multi-agent selection into task bottlenecks and payoff conditions, matching independent subtasks, information coverage and context pollution to different delegation and isolation choices.

Hacker News front page

LLM agents play Pokemon showdown

Oliver Grabner 构建了 Pokémon Showdown 2v2 多智能体环境,Qwen3.8-27B 队友自由交流时对脚本化最高伤害机器人的胜率为64%,禁言时为42%。

AI HOT · Industry

卡兹克详解 AI 辅助发现的 PHP2Shell 漏洞:发现花2天,修复用两个月

卡兹克复盘 GPT-5.6 Sol 辅助发现的 PHP2Shell 漏洞,发现花2天,修复用了两个月。两个远程代码执行漏洞为 CVE-2026-109618 和 CVE-2026-109619,期间发生提前泄密、演示视频流出及多国政府网站预防性下线,补丁于21日晚10点上线。OpenAI 还发布了检测技能。

Hacker News front page

Κ = 0.003 for integer multiplication crossed

整数乘法的条件数 κ 被优化至 0.00300105,较 #4440 提升 4.105e-5,方法为合并查询稀疏化与辅助边界重定时。此前 #4416 版本为 0.00272124,提升 3.477e-5。

Hacker News front page

There isn't an easy exploit for this PHP vulnerability

这处 PHP 漏洞确实存在,但文章判断其利用难度远高于 Log4Shell、Heartbleed 等漏洞,流传的 PoC 无法攻击未经修改的 PHP 实例。文章指出,多份 PoC 依赖本地文件包含(LFI)或人为修改服务端,引入额外漏洞后才展示攻击成功。文章进一步批评将安全判断交给 Claude、GPT 等大语言模型的做法,认为这导致不同开发者重复采用相似的错误利用思路。

TechCrunch · AI

OpenAI and Google's Wiz use AI to find long-hidden PHP flaw, triggering website security alerts

OpenAI and Google-owned Wiz used AI to uncover a serious PHP vulnerability that had gone unnoticed for a long time. The PHP Foundation has released a patch, and the NSA warned sites that cannot patch immediately to take themselves offline for now.

Why it matters: AI found a long-hidden PHP flaw, and as AI lowers the bar for exploitation, patching vulnerable sites becomes more urgent.

Hacker News front page

Omnara launches open-source agent platform as self-hosted alternative to Claude Managed Agents

Omnara is a self-hostable, open-source managed agent platform positioned as an alternative to Claude Managed Agents and OpenAI's Agents API. It works with compatible models and runs on your own machines.

Why it matters: Omnara leaves model choice, the machines agents run on and where agent history is stored to whoever self-hosts it, giving a clearer infrastructure control boundary than vendor-hosted services.

Hacker News front page

CleanFrame: A quieter home screen for Samsung The Frame

开发者发布 CleanFrame,一款面向运行 Tizen 的 Samsung The Frame 电视的简洁主屏,提供应用列表、照片背景和 12 小时制时钟,界面专为电视遥控器设计。

Hacker News front page

Security Disclosure

Railway 今天早上发现客户环境变量存储系统出现异常,调查显示第三方可能访问了加密变量,并取得其中部分变量的解密内容。Railway 已通知上游 AI 推理提供商主动撤销所有密钥,并建议所有用户轮换存储在 Railway 的环境变量、API 密钥及私有 registry 凭据,检查相关服务日志。

The Verge · AI

Meta to create a new class of shares to fund AI push

Meta宣布将发行无投票权的新类别股票,为AI项目、潜在收购和员工薪酬提供资金,计划于2027年开始交易。公司在周三发布的第三季度财报中将今年资本支出预期从$130 billion至$145 billion调整为$135 billion至$145 billion。Meta首席财务官Susan Li表示,新类别股票将扩大融资工具,为AI驱动的业务投资提供灵活性。

The Verge · AI

Data centers could erase decades of air quality gains

Houston Advanced Research Center(HARC)报告预测,若得州所有规划中的数据中心按预期电源结构接入电网,五年内新增污染可能抹去数十年的空气质量改善成果。

OpenAI News

OpenAI extends model access under unified safety standards and updates usage policies

OpenAI is pushing model providers to apply consistent safety standards and has updated its Usage Policies, effective November 20, 2026. Partners must keep safety measures aligned with OpenAI's, investigate potential violations and coordinate enforcement. Services that give direct model access, such as model routers and AI gateways, need review and approval in advance.

Why it matters: OpenAI is extending consistent safety requirements to partners and resale services, and model routers and AI gateways now need approval before offering direct model access.

Hacker News front page

Anthropic opens research preview of Claude Duet, a proactive agent app

Anthropic has opened a research preview of Claude Duet, an app built on Claude Opus 5.5 that can send proactive reminders, carry out tasks and keep working after the user steps away. It draws on the memory, skills and connected apps of the user's Claude account, and needs approval before sending messages, making payments, signing documents or deleting content.

Why it matters: Duet puts background execution and proactive reminders into the workflow while keeping a clear approval boundary on messages, payments and deletions.

TechCrunch · AI

Meta says Muse has over 10M users, with shopping a top activity

Meta在周三第三季度财报电话会上表示,其AI应用Muse用户超过1000万,购物是主要用途之一。Zuckerberg称,约25%完成设置的用户每天使用Muse,在通过智能体购物的活跃用户中,超过四分之一的平均购买金额超过$200。Meta计划对多数用户保持免费,通过交易抽成及高阶模型或更多token的付费使用实现商业化,并预计明年计算成本效率提升超过10x,在收入规模扩大后实现盈利。