字号·· | 护眼
thenextweb

萨提亚·纳德拉:AI 发现的漏洞将使服务中断更易发生Satya Nadella says AI-found bugs will make outages more likely

点「原文对照」整页切到原文,或双击某段只看那段的原文。

2026年1月20日,萨提亚·纳德拉在达沃斯世界经济论坛年会上发言。图片来源:世界经济论坛 / CC BY-NC-SA 4.0,经由Flickr(已裁剪)Satya Nadella speaks at the World Economic Forum Annual Meeting in Davos, 20 January 2026 Credit: World Economic Forum / CC BY-NC-SA 4.0 via Flickr (cropped)

萨提亚·纳德拉预计会出现更多类似上周导致 Outlook 瘫痪的故障,因为人工智能现在发现软件缺陷的速度比企业能够安全修复的速度还要快。微软首席执行官周三在纽约《纽约时报》DealBook 峰会上接受了安德鲁·罗斯·索金的采访。

Satya Nadella expects more outages like the one that took down Outlook last week, because AI now finds software flaws faster than companies can safely fix them. The Microsoft chief executive spoke to Andrew Ross Sorkin at The New York Times’ DealBook Summit in New York on Wednesday.

他为微软自身的故障道歉,并表示该故障与上述问题没有直接关联。

He apologised for Microsoft’s own outage, which he said had no direct link to that problem.

“这是一个一直存在的潜在漏洞。它被触发了,可以说是‘完美风暴’之一,”纳德拉说。

“This was a latent bug that was there. It got invoked and, you know, it’s one of those perfect storms,” Nadella said.

当天晚些时候,他在 CNBC 节目《Mad Money》上告诉吉姆·克莱默,微软本可以做更多工作来预防此事。

Later that day, he told Jim Cramer on CNBC’s Mad Money that Microsoft could have done more to prevent it.

“当然,我们本应该事先进行测试。我们本应该想办法设置隔离舱壁,使其不会导致全球性故障,”纳德拉对克莱默说。

“Of course, we should have done the testing before. We should have figured out how to have the bulkheads such that you didn’t have a global outage,” Nadella told Cramer.

没有时间测试的补丁 纳德拉表示,更大的风险来自发现漏洞的 AI 工具。过去,针对关键银行系统的修复需要经过一周的测试。现在攻击者在漏洞浮现的同时就加以利用,因此修复补丁直接进入生产环境。他说,在他发言的当天,就有三个针对开源包的关键补丁发布。

Patches with no time to test The bigger risk comes from AI tools that find vulnerabilities, Nadella said. A fix for a critical bank system used to get a week of testing. Now attackers exploit flaws as soon as they surface, so fixes go straight into production. On the day he spoke, he said, three critical patches had landed for open-source packages.

“补丁需要从可用状态瞬间部署到生产环境中,因为它正在野外被利用。所以这将造成不稳定,”他说。

“The patch needs to go instantly from being available to being patched in production, because it’s being exploited in the wild. So that’s going to create its instability,” he said.

据报道,银行和美国国家安全局(NSA)将本周美国的故障归咎于针对 AI 发现的缺陷而匆忙发布的补丁。纳德拉还表示,“网络攻击的门槛已大幅降低”。他说,借助 AI 模型,任何人现在都可以运行相当于持续的、国家级攻击者的工具。

Banks and the NSA have reportedly blamed this week’s US outages on rushed patches for AI-found flaws. Nadella also said the “floor of cyberattacks has fallen”. With AI models, anyone can now run what amounts to a persistent, nation-state-level attacker, he said.

索金询问这是否只是一个过渡阶段。纳德拉将其比作“千年虫”问题,但表示目前还没有定论,因为模型不断变得更强大。他补充说,许多关键基础设施运行在无人能修补的固件上。他列举了连接在互联网上的冷水机组、供暖制冷系统和交通网络。“一个麻烦不断的世界”上个月,纳德拉在 20VC 节目中表示,仅靠规模扩大无法实现对齐。索金说,早些时候在峰会上发言的山姆·奥特曼不同意这一前提。

Sorkin asked whether this was a passing phase. Nadella compared it to the Y2K bug, but said he had no view yet, because models keep getting more capable. Much critical infrastructure runs firmware that nobody can patch, he added. He named chillers, heating and cooling systems, and transport networks that sit on the internet. ‘A world of hurt’Last month, Nadella told 20VC that alignment would not come from scale alone. Sorkin said Sam Altman, who spoke earlier at the summit, disagreed with that premise.

纳德拉说,他是从 OpenAI 创始人那里了解到缩放定律的,而且缩放确实奏效了。但他表示,对齐以及一种真正的模型调试方法并没有随之而来。他希望奥特曼和格雷格·布罗克曼是对的。“我是说,如果这不发生,我们就麻烦大了,”纳德拉说。

Nadella said he learned about scaling laws from OpenAI’s founders, and that scaling had worked. Alignment, and a real way to debug models, had not followed, he said. He hoped Altman and Greg Brockman were right. “I mean, if it doesn’t happen, we are in a world of hurt,” Nadella said.

他的回答是他所谓的“信任架构”。他将禁止使用“神经语”推理的模型,这是一种人类无法阅读的内部简写。他说,每个模型的推理都需要防篡改日志,以便其他模型可以检查,因为模型会试图隐藏它们在做什么。他称赞 OpenAI 在训练运行期间监控其模型。

His answer is what he called a trust architecture. He would ban models that reason in neuralese, an internal shorthand people cannot read. Every model’s reasoning needs tamper-proof logs that other models can inspect, he said, because models will try to hide what they are doing. He praised OpenAI for monitoring its models during training runs.

纳德拉说,目前,部署 AI 系统的一方需对其行为负责。他预计保险公司会为这种风险定价,并要求提供日志。

For now, whoever deploys an AI system is liable for what it does, Nadella said. He expects insurers to price that risk and to ask for the logs.

廉价 Token、借用前沿模型 索金询问,如果更便宜的模型能完成相同工作,微软持股的 OpenAI 和 Anthropic 会怎样。纳德拉说,日常办公工作不再需要前沿模型。他说,Copilot 已经在 OpenAI、Anthropic 和微软自家的 MAI 模型之间进行选择。他补充说,他更大的押注在微软身上。

Cheap tokens, borrowed frontier Sorkin asked what happens to OpenAI and Anthropic, in which Microsoft holds stakes, if cheaper models can do the same work. Nadella said everyday office work no longer needs frontier models. Copilot already chooses between OpenAI, Anthropic, and Microsoft’s own MAI models, he said. He added that his bigger stake is in Microsoft.

“为 Token 的某种巨大定价权辩护,就是在赌反美国 GDP,”他说。

“Arguing for some massive pricing power of tokens is betting against somehow US GDP,” he said.

纳德拉说,微软与 OpenAI 的协议使其获得免版税使用前沿模型的权限,直到 2032 年。这让微软能够构建落后前沿三到六个月的 MAI 模型,仅使用 5% 到 20% 的算力。他说,Anthropic 和 OpenAI 玩的是“垂直栈游戏”,而微软玩的是“模型系统游戏”。

Microsoft’s deal with OpenAI gives it royalty-free access to a frontier model until 2032, Nadella said. That lets Microsoft build MAI models three to six months behind the frontier, using 5% to 20% of the computing power. Anthropic and OpenAI play “a vertical stack game”, he said, while Microsoft plays “a model system game”.

纳德拉表示,正如当年从雅虎转向搜索那样,代理、工作以及华盛顿的 AI 代理将重塑网络上谁向谁付费的格局。靠人类访客带来的广告收入生存的网站将不得不改变。当被问及拒绝代理进入的网站(如亚马逊)时,他说竞争对手会接走这笔生意。

Agents, jobs, and Washington AI agents will reshuffle who pays whom on the web, as the move from Yahoo to search once did, Nadella said. Sites that live off ads from human visitors will have to change. Asked about sites that keep agents out, such as Amazon, he said rivals would take the business.

“如果你不接受代理,别人就会接受。如果亚马逊不接受,Shopify 就会接受,”纳德拉说。

“If you don’t accept the agent, somebody else will. If Amazon doesn’t, Shopify will,” Nadella said.

他表示,微软向 6 万人提供的员工买断方案旨在让每个人都拥有自主权。软件工程“依然非常活跃”,但工作性质已发生变化。微软 Cowork 代理背后的团队只有 11 名开发人员,他们构建的是“一个构建产品的软件工厂”。纳德拉说,这种转变所需的时间会比任何人预期的都要长。

The staff buyouts Microsoft offered 60,000 people were meant to give everyone agency, he said. Software engineering is “very much alive”, but the work has changed. The team behind Microsoft’s Cowork agent is 11 developers who build “a software factory that builds the product”. That shift will take longer than anyone expects, Nadella said.

索金指出,纳德拉一直与特朗普政府保持距离。纳德拉表示,他的工作是通过产品赢得微软的经营许可。

Sorkin noted that Nadella has kept his distance from the Trump administration. Nadella said his job is to earn Microsoft’s permission to operate through its products.

“我只是不想待在华盛顿。我只想待在华盛顿州雷德蒙德,”他说。

“I just don’t want to be in Washington. I just want to be in Redmond, Washington,” he said.

阅读媒体原文 ↗