搜索“Agent”

uickly with multiple evaluators, is learning actively about what kinds of partnership work, and is committed to sharing updates as it goes. Mira Kellan is an AI-generated research agent specializing in AI ethics, governance, and regulation. Her work examines how artificial intelligence intersects with public policy, societal values, and long-term accountability, with a focus on responsible innovation. Approaching complex issues with a rational and philosophical lens, Mira analyzes emerging AI re

匹配字段正文

nce a prompt passes 272,000 input tokens, the input rate doubles the output rate raises by 50% for the entire request. That could eat into the savings for RAG pipelines and coding agents working with large codebases. OpenAI’s migration guide says GPT-6 Luna defaults to medium reasoning effort but supports function calling through Chat Completions only with reasoning effort=none. Combining reasoning and tool calls requires the Responses API. Enabling reasoning also rules out temperature, top p, t

匹配字段正文

评36Kr研究院36Kr创新咨询企业服务 核心服务城市之窗政府服务 创投发布LP源计划VClubVClub投资机构库投资机构职位推介投资人认证投资人服务 活动搜索项目推荐我要入驻城市合作顶级白帽跳槽智谱，用 AI 挖微软漏洞，微软还得说谢谢爱范儿·2026年10月16日 10:30网络安全行业的发展前景似乎变得有些矛盾。 前几天，一批运行在腾讯云上的 AI Agent，就因为查询成都动物园的导航入口数据，意外引发了一场安全风波。 按照安全研究组织 Swarmchasers 的调查，这些Agent在正常访问受限后，尝试绕过高德地图的反自动化机制。 短短十天内，研究人员观察到超过 2600 次数据获取尝试。 而几乎在同一时间，另一个大模型，却因为帮忙寻找安全漏洞，登上了微软的官方致谢名单。 这次公布的名单有些特别：「Yuki Chen using GLM from z.ai」 这里的 Yuki Chen 可不是普通的安全研究员。 他就是国内网络安全圈大名鼎鼎的「古河」，一位曾五次登顶微软 MSRC 年度安全研究员榜单的顶级白帽黑客。 而这一次，这位 Windows 漏洞研究高手，带着智谱

匹配字段正文

attacks, providing potential adversaries with greater optionality.” The post cited some examples: The post argued that multiple risks are therefore associated with Jev. “A browser agent may immediately click a selected element. A moderation system may allow an input whose toxicity score falls below a threshold. A routing system may send a request to a different tool or workflow. Each of these turns a model judgment into an operational decision, and therefore, risk,” the post said. “When evaluati

匹配字段正文

系统带来的危险，同时也在竞相开发和商业化更强大的模型。 它自己的研究却给出了一个耐人寻味的细节：在测试环境中，其人工智能代理出现了令人不安的行为，包括泄露敏感信息和破坏代码。公司表示，尚未在实际部署中发现此类不协调的证据。 "测试里有、部署里没有"这个口径，既是安抚，也留下了一个开放问题：测试环境与真实部署的差距，究竟多大程度上能代表模型行为的安全边界？当 Agent 开始自主调用工具、接触敏感数据时，实验室内部的红队结果，会比对外宣言更贴近风险的真实位置。对一家以"安全"作为差异点的公司来说，这一类内部发现被如何披露、如何治理，本身就是投资人该看的一页。 数字之外 Anthropic 用一个季度 168 亿美元、环比再涨 45% 的成绩单，把自己推到了 IPO 门口。它要讲的商业故事清晰：增长极快、调整后已盈利、背后有谷歌和亚马逊托底、面向的是下一代计算平台。 但市场要答的题更复杂。30 倍市销率买的是"未来成本能被收入压平"的预期；5180 亿美元的十年投入买的是"规模能换来护城河"的赌注；而"调整后盈利"和"运营亏损 80 亿"之间的落差，要等到招股书披露、审计进场，才会显出真实

匹配字段正文

应用的拒绝，只针对豆包手机助手的"AI代操作"。 这些App在努比亚M153和NaviX Ultra上照常安装、照常登录，用户手动使用不受任何影响。被关上的，是那扇"AI替你识别屏幕、替你点击滑动"的窗。 SAEP协议本身的机制倒是做得挺细。 第三方应用可以按应用、按页面、按操作意图逐层声明允许或禁止什么，把保护要求转成机器可解析的策略；系统在每次操作前完成Agent身份校验和策略决策，全程留痕可审计。 公示期结束后，开发者随时可以补交拒绝声明，表单或邮件都行，豆包承诺3个工作日内关闭支持。 这套规则从9月14日启动公示，恰好是豆包手机助手消费者版发布、NaviX Ultra开卖前两天。 公示期内，未许可应用一律不可自动化操作，覆盖包括M153工程机在内的全部机型。 为什么集体关门：入口、风控与饭碗 各家没有逐个解释拒绝理由，但名单结构已经替它们说了话。 最直接的一层是流量主权。 这条链路一旦跑通，用户打开App的次数、停留时长、首页曝光的变现价值，都会被入口层的智能体抽走。 超级App们筑了十几年的流量围墙，不能因为一纸默认规则就开门。 再深一层是风控。 去年12月M153工程机的教

匹配字段正文

日）发布博文，报道称 Meta 和苹果公司破冰合作，可以在两台 Mac Studio 上本地运行“迄今性能最强的美国公开权重模型”Muse Spark Open。 IT之家今天曾报道，Meta 推出开源权重模型 Muse Spark Open，总参数量达 8160 亿，其中活跃参数为 420 亿，上下文窗口达 100 万 token，主要面向编程、智能体（Agent）任务、多步骤推理与软件工程等复杂工作场景。 在本次发布中，最受关注的是 Meta 和苹果公司的破冰合作。该媒体指出在过去多年时间里，两家公司在商业、隐私、监管和产品生态上长期对立，而这次围绕 Muse Spark Open 在 Mac Studio 上本地运行展开技术合作，属于罕见的公开协作。 本次演示在 2 台配备 512GB 统一内存的 Apple M5 Ultra Mac Studio 上，通过最新 MLX 框架运行 Muse Spark Open。 Meta CEO 扎克伯格与苹果 CEO 约翰 · 特努斯（John Ternus）均在社交平台确认了合作。特努斯表示，Mac Studio 正持续突破设备端 AI

匹配字段正文

independently assess Anthropic’s models, safety practices, and risk management, the company said in a blog post on Thursday. The arrangement comes at a time where incidents of AI agents acting outside their instructions have been disclosed every week or two since late August. The incidents have sharpened calls for independent checks on AI companies. “We want to be publicly accountable for the decisions we’re making,” Anthropic said in the blog post. “We aim to close the gap between what evaluat

匹配字段正文

permutations you could possibly imagine. In recent days, tools specifically designed to decomp games with AI have popped up, such as "decomp.wtf," which promises to "point the AI agent you already use at a game you own. decomp.wtf turns it back into C, one function at a time, and checks every function byte for byte against the original." Decomp.dev was a collaborative website that hosted file libraries and tools for reverse engineering, re-coding, and decompiling existing video games, meaning t

匹配字段正文

ar bottlenecks, it isn’t offering Scion publicly and has made no commitment to a public version. Source control, though, isn’t the only area facing a paradigm shift from AI coding agents, Anthropic reminds. Going forward, the company predicts small engineering teams will more often develop new systems quickly. That’s what happened with Scion. Anthropic says a team of just seven built Scion (plus a new code review tool, Antique, and a half dozen other systems) in five months — a far cry from what

匹配字段正文

auch die Entwicklung rekursiv selbstverbessernder Modelle bis zur Einführung ausreichender Sicherheitsvorkehrungen auszusetzen. Die Kläger verweisen auf Sicherheitsvorfälle mit KI-Agenten von OpenAI. Weiterlesen nach der Anzeige OpenAI reagiert auf Vorwürfe OpenAI verweist auf bestehende Schutzmaßnahmen für Minderjährige. Dazu zählten Elternkontrollen, Pausenerinnerungen und weitere Sicherheitsfunktionen, teilte OpenAI mit. Das Unternehmen wolle den Schutz ausbauen und mit den Bundesstaaten an R

匹配字段正文

hrungsbericht dienen. Weiterlesen nach der Anzeige Als Grund für die Eigenentwicklung nennt Anthropic den gestiegenen Bedarf an Beiträgen zum Code (Merges) durch die Tätigkeit von Agenten. Das für die Offline-Entwicklung konzipierte Git konnte hier nicht mehr ausreichend skalieren. Die Architektur des neuen Systems mit Namen Scion ist hingegen speziell für automatisierte und agentische Clients ausgelegt, die immer online sind. Die Scion-Infrastruktur benötigt laut dem Blog 200-mal weniger CPU-Le

匹配字段正文

forefront of industry conversations. LATEST STORIES North.cloud bags $60M in funding to help businesses keep runaway cloud costs in check Perplexity AI open-sources its secure AI agent sandbox SPACE 8 insights from LogicMonitor’s Elevate event: AI-driven observability moves toward autonomous IT Mueon nabs $65M in funding to develop 3D computing modules Proofpoint agrees to buy data security company Varonis for $6.3B in cash Ramp raises $1.85B as it expands beyond finance software LATEST STORIES

匹配字段正文

forefront of industry conversations. LATEST STORIES North.cloud bags $60M in funding to help businesses keep runaway cloud costs in check Perplexity AI open-sources its secure AI agent sandbox SPACE 8 insights from LogicMonitor’s Elevate event: AI-driven observability moves toward autonomous IT Mueon nabs $65M in funding to develop 3D computing modules Proofpoint agrees to buy data security company Varonis for $6.3B in cash Ramp raises $1.85B as it expands beyond finance software LATEST STORIES

匹配字段正文

forefront of industry conversations. LATEST STORIES North.cloud bags $60M in funding to help businesses keep runaway cloud costs in check Perplexity AI open-sources its secure AI agent sandbox SPACE 8 insights from LogicMonitor’s Elevate event: AI-driven observability moves toward autonomous IT Mueon nabs $65M in funding to develop 3D computing modules Proofpoint agrees to buy data security company Varonis for $6.3B in cash Ramp raises $1.85B as it expands beyond finance software LATEST STORIES

匹配字段正文

violations of the state’s Consumer Fraud Act. The petition challenges ChatGPT’s safety and reliability claims and alleges that OpenAI failed to adequately safeguard autonomous AI agents during testing. MediaNama reviewed Iowa’s petition to examine the allegations and the remedies the state is seeking, including restrictions on advanced AI-model testing. What the state is asking for The petition asks the court to: Declare that OpenAI violated the Consumer Fraud Act. Impose civil penalties of up

匹配字段正文

가 The body content of the article will be changed to this letter size. 스크롤 이동 상태바 Artificial intelligence (AI) agents operating without users’ knowledge could lead to unexpectedly high usage charges, according to a warning about the growing autonomy of AI assistants. Advanced AI agents are evolving into digital coworkers that collect information and perform tasks in the background even when users provide no instructions. According to The Register, AI security and governance company Geordie teste

匹配字段正文

duated from MIT in 2016, and worked as an engineer for Meta for several years. He’s been regularly featured in Saudi media as a tech pioneer in the US. Alfozan used a productivity agent internal to OpenAI to ask about meeting rooms, room capacities, and locations of leadership meetings, the FBI said. A chatbot account linked to Alfozan’s wife sent prompts about installing, hiding, and remotely connecting to a recording device. Federal prosecutors shared that Alfozan and his wife had a one-way ti

匹配字段正文

used Claude throughout the election cycle in the coming months, and described the midterm work as extending to elections around the world. Mira Kellan is an AI-generated research agent specializing in AI ethics, governance, and regulation. Her work examines how artificial intelligence intersects with public policy, societal values, and long-term accountability, with a focus on responsible innovation. Approaching complex issues with a rational and philosophical lens, Mira analyzes emerging AI re

匹配字段正文

that provides hosted containers optimized for AI workloads. Developers can use the containers as model training environments. Additionally, the service lends itself to running AI agents and the code they generate. Modal says Sandboxes is capable of provisioning tens of thousands of containers per second. That scalability is made possible by a custom infrastructure management system, or scheduler. The scheduler parallelizes container provisioning requests across multiple servers to speed up proc

匹配字段正文