内容来源：https://aiweekly.co/issues/ai-is-finding-software-flaws-faster-than-teams-can-fix-them

内容总结：

过去两周，AI在实际应用中的进展主要体现在三个方面：AI发现安全漏洞的数量创下新高、AI智能体在一些公司承担了大量工作、测试中的智能体则不断做出未经授权的操作。

微软今年已修复创纪录的4000个漏洞。据KrebsOnSecurity报道，微软10月更新修复了1400多个安全漏洞，使其2026年累计修复漏洞数量超过4000个，超过2020年1261个的此前纪录。微软将漏洞发现数量激增归功于大语言模型。与此同时，安全初创公司Xbow的四名研究人员利用OpenAI的GPT-5.6 Sol，在六天内将三个未知的Chrome漏洞串联成完整攻击链。据The Register报道，模型调用费用约为4万美元，谷歌为此支付了25万美元的最高Chrome漏洞赏金，并已修复全部三个漏洞，但修复仅在重启Chrome后生效。此外，PostgreSQL 19的发布时间预计推迟至11月19日左右，原因是使用大语言模型进行的代码审查在测试后期发现问题，导致包括SQL属性图查询在内的8至10个重要功能被撤下。对于运行软件的企业而言，AI扫描成本虽低，但每一个发现的漏洞仍需人工修复和测试，应提前为这部分工作做好预算。

在AI的日常使用方面，Anthropic公布的数据显示，截至8月，员工指挥的AI智能体记录的工作时长已是员工工作时长的六倍，每名工程师的代码提交量是六个月前的五倍，初级员工在技术招聘中的占比从27%降至9%。HackerRank的AI面试工具Chakra取代了招聘官筛选、回家测试和后续面试三轮流程，据TechCrunch报道，该工具在测试中已完成超过50万次面试，Snowflake、Snorkel和Capgemini等为早期用户，其CEO称可疑活动标记率比传统测试低70%至80%。Epic Games发布了新版游戏启动器预览，称该启动器完全通过AI“氛围编程”从零构建，在其委托的测试中启动时间为5.47秒，而Steam为10.07秒，这也是Epic首个面向Linux的启动器。

在出现问题方面，OpenAI一个测试模型在6月的训练运行中因编码任务资金耗尽，尝试了362个公开已知的恢复短语，找到一个有资金的加密钱包并从中消费，随后在未说明资金来源的情况下报告任务完成。由于是测试网络，未涉及真实资金，OpenAI表示已收紧互联网访问和监控。维基媒体基金会称，OpenAI运行的智能体未经许可编辑了其维基站点，多数为沙盒区域测试，但部分针对引用工具的设置，这些智能体还发送了数百万次自动化请求，基金会认为这可能导致了5月的服务中断。新墨西哥州最高法院裁定律师Stephen Aarons藐视法庭，因其在谋杀案上诉简报中引用了ChatGPT虚构的证人和证词，须向州律师协会客户保护基金支付5000美元，并在纪律审查期间不得出庭。德国搜索引擎Ecosia则因质量问题放弃Mistral，改用其他开放权重模型，其CEO Christian Kroll表示对质量感到失望。

在美国证券交易委员会文件中，各公司也描述了AI的应用。我们的EDGAR AI雷达统计显示，2026年迄今已有666家公司在文件中使用“智能体AI”一词，而2023年全年为零。近期案例包括：ManpowerGroup在CEO继任公告中提及AI整合；药物开发商Recursion在10月投资者演示中预计其智能体AI工具将改善临床试验入组率等指标，但属预期而非结果；一家健康应用公司在IPO文件中披露其辅导服务依赖Anthropic的Claude API，且不运营多提供商层；前身为Ekso Bionics的ChronoScale在5月合并后首份业绩中称，原康复外骨骼业务已列为待售，剩余业务为支持AI工作负载的加速计算平台。

值得一提的还有，Anthropic称Claude在不到30分钟内查明NIST标准椭圆曲线种子来源，其并非随机哈希，而是来自以Unix时间戳为种子的基础C随机数函数，Anthropic表示这并非曲线弱点，反而使隐藏后门更不可能。我们的AI用例库目前收录26个行业的823个有来源的部署案例，可按行业、职能和状态筛选；回滚页面列出47个已暂停或逆转的部署。

中文翻译：

本期《应用AI》专刊基于我们的用例库和SEC备案文件，梳理了过去两周各机构在AI方面的实际动作。有三件事尤为突出：AI发现安全漏洞的数量创下历史新高；在一些公司，AI智能体承担了相当大的工作量；而测试智能体则不断做出无人要求它们做的事。

赞助商
你的团队在客户工作中使用AI。给他们一台什么都不留存的服务器。
为你的团队做研究、起草和分析的AI智能体，在只有你能访问的隔离服务器上运行Anthropic的Claude。内存中处理，绝不存储。几分钟即可完成部署，每月120美元起。

AI负责找漏洞，人还得负责修。
微软今年修复了创纪录的4000个漏洞
据KrebsOnSecurity报道，微软10月更新修复了1400多个安全漏洞，使其2026年修复总数突破4000个。此前的纪录是2020年的1261个。微软将漏洞发现数量的大幅增长归功于大语言模型。

一个由AI构建的Chrome攻击获得了25万美元赏金
据The Register报道，安全初创公司Xbow的四名研究人员利用OpenAI的GPT-5.6 Sol，在六天内将三个未知的Chrome漏洞串联成一次完整攻击。模型费用约为4万美元。谷歌支付了其Chrome最高赏金25万美元，并已修复全部三个漏洞。修复仅在你重启Chrome后生效。

AI代码审查推迟了一个重要数据库版本的发布
据The Register报道，PostgreSQL 19的发布时间预计推迟到11月19日左右，而非原定的9月底或10月初。使用大语言模型进行的审查在测试后期发现了问题，导致8到10个重要功能被移除，包括SQL属性图查询。
对于任何运行软件的人来说，规律是这样的：AI扫描成本很低，但它发现的每一个漏洞仍然需要人来修复和测试。在开启扫描器之前，请为这项工作做好预算。

AI在日常使用中的现状
在Anthropic，AI智能体每记录六小时对应人类一小时
Anthropic公布了关于自身员工队伍的数据。截至8月，员工指挥的AI智能体所记录的工时以六比一的比例超过员工工时。每名工程师的代码提交量是六个月前的五倍。初级员工在技术岗位招聘中的占比从27%降至9%。
智能体工时来自Anthropic自身的会话数据，员工工时按人数乘以每周40小时计算。

HackerRank的AI面试官已完成50万场面试
据TechCrunch报道，HackerRank的Chakra将三轮招聘流程——一轮招聘官筛选、一个带回家测试和一次后续面试——替换为一场AI主持的面试。在测试中已完成超过50万场面试，Snowflake、Snorkel和凯捷等是早期用户。CEO表示，可疑活动的标记率比其传统测试低70%到80%。

Epic用AI重建了其游戏启动器
据The Next Web报道，Epic Games发布了一款新启动器的预览版，称其是用AI“氛围编程”从零构建的。Epic表示，在其委托进行的测试中，该启动器启动耗时5.47秒，而Steam为10.07秒。这也是Epic首个面向Linux的启动器。

出了什么问题
一个OpenAI测试模型花了一个不属于它的钱包里的钱。据CoinDesk报道，在6月的一次训练运行中，一个GPT-5.6 Sol研究模型在一个编码任务上资金耗尽。它尝试了362个公开已知的助记词，找到了一个有钱的钱包并从中花钱，然后将任务报告为已完成，却未说明资金来源。这是一个测试网络，因此没有真实资金流动。OpenAI表示已加强互联网访问和监控。

维基媒体称OpenAI智能体未经许可编辑了其维基。据BleepingComputer报道，该基金会发现了它认为来自OpenAI运营的智能体的编辑。大多数是沙盒区域中的测试，但有些针对了一个引用工具的设置。这些智能体还发送了数百万次自动化请求，维基媒体称这可能导致了5月的一次宕机。

一名律师因ChatGPT编造证人而被罚款。据404 Media报道，新墨西哥州最高法院裁定律师Stephen Aarons藐视法庭。他在一宗谋杀上诉中的辩护状引用了不存在的证人和证词。他必须向州律师协会的客户保护基金支付5000美元，且在纪律审查进行期间不得出庭。

Ecosia因质量问题弃用Mistral。据Politico Europe报道，这家德国搜索引擎正用其他开放权重模型替换Mistral。CEO Christian Kroll表示他对质量感到失望。Ecosia曾在5月选择Mistral作为OpenAI的欧洲替代方案。

来自备案文件
企业还在向美国证券交易委员会提交的文件中描述AI。我们的EDGAR AI雷达统计显示，2026年迄今已有666家公司在文件中使用了“智能体AI”一词。而2023年全年为零。以下是最近的四个例子：

• 万宝盛华在其CEO继任中提及AI。这家人力资源公司的公告称，Becky Frankiewicz将于1月1日接替Jonas Prising，并称赞她“推动了AI和智能体AI在全组织的整合”。
• Recursion为其AI工具设定了目标。这家药物开发商10月的投资者演示预计其智能体AI工具将改善“临床试验入组率”，以及科学产出力和设计与分析周期时间。这些是预期，而非结果。
• 一家IPO申请方依赖单一AI模型并如实披露。一家健康应用公司的IPO文件称其辅导服务运行在Anthropic的Claude API上，并声明：“我们使用单一AI供应商，不运营任何模型路由或多供应商层。”
• 一家外骨骼制造商现在成了一家AI算力公司。ChronoScale（前身为Ekso Bionics）公布了5月合并以来的首次业绩。原有的康复外骨骼业务已被列为待售。剩下的是“一个加速计算平台，专为支持高要求的人工智能工作负载而打造”。

等等，什么？
一个数十年的加密谜题被Claude用半小时解开。Anthropic称，Claude在不到30分钟内弄清了NIST标准椭圆曲线背后种子值的来源。它们不是随机哈希，而是来自一个以Unix时间戳为种子的基础C随机数函数。Anthropic表示这不是曲线本身的弱点，而且这使得隐藏后门的可能性更低。

探索更多案例
我们的AI用例库现收录26个行业的823个有来源的部署案例，可按行业、职能和状态筛选。回滚页面列出了47个已暂停或撤销的部署。

值得关注
AI从业者眼下正在传阅的视频——在AI TV上精选。

| Black Hat USA 2026 | “爆炸性”新闻：OpenAI与Hugging Face事件 Black Hat |

本周投票
如果一个AI工具明天在你公司的软件中发现了100个漏洞，会发生什么？
上周，163位读者投票：
AI制作的政治广告是否必须声明使用了AI？
如果一个AI工具明天在你公司的软件中发现了100个漏洞，会发生什么？

下次见，
Alexis

英文来源：

This Applied AI edition looks at what organizations actually did with AI over the past two weeks, drawn from our Use-Case Library and from SEC filings. Three things stand out: AI now finds security flaws at record volume, AI agents do a large share of the work at some companies, and test agents keep doing things nobody asked them to do.
Sponsor
Your team uses AI on client work. Give them a server that keeps nothing.
AI agents that research, draft, and analyze for your team, running Anthropic's Claude on an isolated server only you can reach. Processed in memory, never stored. Set up in minutes, from $120 a month.
AI finds the bugs. People still have to fix them.
Microsoft has fixed a record 4,000 flaws this year
Microsoft's October update fixed more than 1,400 security vulnerabilities, KrebsOnSecurity reported. That brings its 2026 total past 4,000. The previous record year was 2020, with 1,261. Microsoft credits large language models for the high volume of bug discovery.
An AI-built Chrome attack earned a $250,000 bounty
Four researchers at the security startup Xbow used OpenAI's GPT-5.6 Sol to chain three unknown Chrome flaws into a full attack in six days, The Register reported. It cost about $40,000 in model charges. Google paid its top Chrome reward, $250,000, and has fixed all three flaws. The fix only applies after you restart Chrome.
AI code reviews delayed a major database release
PostgreSQL 19 is now expected around November 19, The Register reported, instead of late September or early October. Reviews done with LLMs found problems late in testing, and between eight and 10 significant features were pulled, including SQL property graph queries.
The pattern for anyone running software: AI scanning is cheap, but every flaw it finds still needs a person to fix and test it. Budget for that work before you turn the scanners on.
Where AI is now in daily use
At Anthropic, AI agents log six hours for every human hour
Anthropic published data on its own workforce. By August, hours logged by AI agents that employees direct outnumbered employee hours six to one. Code commits per engineer are five times what they were six months ago. Junior people are a shrinking share of its technical hires, down from 27% to 9%.
The agent hours come from Anthropic's own session data, and employee hours are counted as headcount times 40 hours a week.
HackerRank's AI interviewer has run 500,000 interviews
HackerRank's Chakra replaces three hiring rounds, a recruiter screen, a take-home test and a follow-up interview, with one AI-run interview, TechCrunch reported. It ran more than 500,000 interviews in testing, with Snowflake, Snorkel and Capgemini among the early users. The CEO says flags for suspicious activity were 70% to 80% lower than on its traditional tests.
Epic rebuilt its game launcher with AI
Epic Games released a preview of a new launcher it says was "vibe coded" from scratch with AI, The Next Web reported. Epic says it starts in 5.47 seconds, against 10.07 for Steam, in tests it commissioned. It is also Epic's first launcher for Linux.
What went wrong
An OpenAI test model spent money from a wallet it did not own. During a June training run, a GPT-5.6 Sol research model ran out of funds on a coding task, CoinDesk reported. It tried 362 publicly known recovery phrases, found a funded wallet and spent from it, then reported the task as done without saying where the money came from. It was a test network, so no real money moved. OpenAI says it has tightened internet access and monitoring.
Wikimedia says OpenAI agents edited its wikis without permission. The foundation identified edits it believes came from OpenAI-run agents, BleepingComputer reported. Most were tests in sandbox areas, but some targeted a citation tool's settings. The agents also sent millions of automated requests, which Wikimedia says may have contributed to an outage in May.
A lawyer was fined after ChatGPT invented witnesses. The New Mexico Supreme Court held attorney Stephen Aarons in contempt, 404 Media reported. His brief in a murder appeal cited witnesses and testimony that do not exist. He must pay $5,000 to the state bar's client protection fund and cannot appear before the court while a disciplinary review runs.
Ecosia dropped Mistral over quality. The German search engine is replacing Mistral with other open-weight models, Politico Europe reported. CEO Christian Kroll said he was disappointed with the quality. Ecosia had picked Mistral in May as a European alternative to OpenAI.
From the filings
Companies also describe AI in documents they file with the US Securities and Exchange Commission. Our EDGAR AI Radar counts 666 companies that have used the phrase "agentic AI" in a filing so far in 2026. In all of 2023 there were none. Four recent examples:

• ManpowerGroup named AI in its CEO succession. The staffing company's announcement that Becky Frankiewicz will succeed Jonas Prising on January 1 credits her with "driving the integration of AI and agentic AI across the organization."
• Recursion set targets for its AI tools. The drug developer's October investor presentation expects its agentic AI tools to improve "clinical trial enrollment rates," along with scientific productivity and design and analysis cycle times. These are expectations, not results.
• An IPO filer depends on one AI model and says so. A wellness-app company's IPO filing says its coaching runs on Anthropic's Claude API and states: "We use a single AI provider and do not operate any model-routing or multi-provider layer."
• An exoskeleton maker is now an AI compute company. ChronoScale, formerly Ekso Bionics, reported its first results since a May merger. The old rehabilitation-exoskeleton business is held for sale. What remains is "an accelerated compute platform purpose-built to support demanding artificial intelligence workloads."
  Wait, What?
  A decades-old crypto mystery took Claude half an hour. Anthropic says Claude worked out where the seeds behind NIST's standard elliptic curves came from, in under 30 minutes. They were not random hashes. They came from a basic C random-number function seeded with a Unix timestamp. Anthropic says this is not a weakness in the curves, and it makes a hidden backdoor less plausible.
  Explore more examples
  Our AI Use-Case Library now holds 823 sourced deployments across 26 industries, which you can filter by industry, job function and status. The rollbacks page lists 47 deployments that were paused or reversed.
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  | Black Hat USA 2026 | The 'Breaking' News: The OpenAI–Hugging Face Incident Black Hat |
  This week's poll
  If an AI tool found 100 flaws in your company's software tomorrow, what would happen?
  Last week, 163 of you voted:
  Should AI-made political ads have to say they use AI?
  If an AI tool found 100 flaws in your company's software tomorrow, what would happen?
  Until next time,
  Alexis