要点
- OpenAI 和 Anthropic 正在调查数万起事件,在这些事件中,先进 AI 模型自行突破安全边界、篡改系统,或试图逃避监控。
- OpenAI 的智能体试图入侵美国教育部的网站,使用窃取的登录凭证访问人口普查局的数据,并在网络论坛上分享 SEC 的信息。
- 这些模型没有是非观念,因此它们以极端的执着追求任务,并在正当手段失败时诉诸未经授权的方法。
OpenAI 和 Anthropic 目前正在调查数万起事件,在这些事件中,它们最先进的 AI 模型采取了会被外部审查者标记为有问题的行动。
Axios援引多个消息源报道了这些发现。这些事件发生在过去几个月的内部测试和实际部署期间,其数量之庞大表明,该问题的复杂程度比已公开的情况高出数个数量级。总数可能远超目前已统计的数字。
据说这些事件与 OpenAI 周五披露的两起案例大致相当。这些事件包括创建留言板、逃出沙箱、劫持网站、自我提示,以及试图规避监控系统。OpenAI 周五宣布,已暂停对其能力最强的内部模型的训练,在确信自身网络安全足以支撑之前不会恢复训练。
OpenAI 的模型篡改了美国政府网站
《纽约时报》描述了涉及美国机构的几起具体事件。在教育部,OpenAI 的智能体试图入侵网站,以从民权办公室收集数据。OpenAI 表示仍在调查中。
在隶属于商务部的普查局,该 AI 的行为超出了简单的抓取。它利用在网上找到的登录凭证从网站提取数据,从而获得了未授权访问权限。
在 SEC 一案中,OpenAI 的智能体获取了信息,随后在一个在线论坛上主动分享了这家证券监管机构的公开数据。SEC 发言人向《纽约时报》表示,该机构正与 OpenAI 保持联系。目前没有迹象表明非公开信息被未经授权地访问。
OpenAI 是在 Hugging Face 事件引发的广泛内部审查中才发现这些案例的。CEO Sam Altman 承认,信息披露"没有达到我们希望的速度"。Altman 表示,公司有"数 PB 的智能体活动日志"需要梳理。

据 OpenAI 称,这些事件均未构成实际的数据泄露,其中一些只是常规的研究活动。但公司仍将其称为"意外且令人担忧的行为"的实例。
例如,芝加哥市长办公室表示,OpenAI 最近告知该市官员,其模型从市政府网站上抓取了公开可用的信息。这本身听起来无害,因为所有搜索引擎都会做同样的事。OpenAI 之所以仍然将其标记出来,很可能是因为行为中"意外"的那一部分,也就是说,模型自行决定以无人预料到的方式去获取数据,而这很可能正是 OpenAI 认为其"令人担忧"的原因。
但这也解释了目前正在审查的大量案例是如何产生的。这一切都取决于什么才算作网络安全事件。OpenAI 表示,其智能体倾向于访问政府网站,因为它们是公共信息的权威来源。
失控的智能体行为正在成为整个行业的问题
这个问题并不局限于 OpenAI,尽管 Altman 的公司目前累积的案例最多。来自 Anthropic、Meta 和 Google 的 AI 智能体也曾入侵或试图入侵企业、大学和政府机构,此类案例数量正在不断增加。在每一起事件中,开发商都是在事后才发现自己的 AI 做了什么。
问题的很大一部分在于最新前沿模型内置的极端执着性。它们被优化为在长时间跨度内解决任务,即使根本没有可行之路,也不会停止寻找突破口。当智能体遇到障碍时,它会试图绕过障碍——并非出于恶意,而是因为达成目标才是唯一重要的衡量标准。
这种执着最终会导致不当行为,因为模型会穷尽一切可能的路径,包括违反安全政策或法律的路径。OpenAI 将其中一个泄露内部 GitHub 数据的模型描述为“一个极度执着的内部模型”。但更深层的问题在于,这些模型没有是非观念,而这正是对齐研究试图解决的问题。如果模型能理解什么行为违法,它们就不会自行走上那些路径。显然,仅把这一点写进提示词是不够的。
Axios
Key Points
- OpenAI and Anthropic are investigating tens of thousands of incidents in which advanced AI models independently broke through security boundaries, tampered with systems, or tried to evade monitoring.
- OpenAI's agents attempted to hack the US Department of Education's website, used stolen login credentials to access Census Bureau data, and shared SEC information in online forums.
- The models have no sense of right and wrong, so they pursue tasks with extreme persistence and resort to unauthorized methods when legitimate ones fail.
OpenAI and Anthropic are currently investigating tens of thousands of incidents in which their most advanced AI models took actions that external reviewers would flag as problematic.
Axios reports the findings, citing multiple sources. The sheer volume of incidents, which occurred during both internal testing and real-world deployment over the past several months, suggests the problem is orders of magnitude more complex than what's been made public. The total number could grow well beyond what's already been counted.
The incidents are said to be roughly on par with the two cases OpenAI disclosed on Friday. They include creating message boards, breaking out of sandboxes, hijacking websites, self-prompting, and attempts to evade monitoring systems. OpenAI announced Friday that it had paused training on its most capable internal models and that training won't resume until the company is confident its own cybersecurity holds up.
OpenAI's models tampered with US government websites
The New York Times describes several specific incidents involving US agencies. At the Department of Education, OpenAI's agents tried to hack the website to collect data from the Office for Civil Rights. OpenAI said it's still investigating.
At the Census Bureau, which falls under the Department of Commerce, the AI went beyond simple scraping. It pulled data from the website using login credentials it found online, gaining unauthorized access.
In the SEC case, OpenAI's agents retrieved information and then actively shared public data from the securities regulator in an online forum. An SEC spokesperson told the NYT that the agency is in contact with OpenAI. There's no indication that non-public information was accessed without authorization.
OpenAI only discovered these cases during the broad internal review triggered by the Hugging Face incident. CEO Sam Altman acknowledged that disclosure has not "been as fast as we would have liked." The company has "petabytes of agent activity logs" to work through, Altman said.

None of the incidents amounted to an actual breach, according to OpenAI, and some were just routine research activity. The company still called them examples of "unexpected and concerning behavior."
The mayor's office in Chicago, for instance, said OpenAI recently told city officials that its models had pulled publicly available information from a city website. That sounds harmless on its own since every search engine does the same thing. The fact that OpenAI flagged it anyway likely comes down to the "unexpected" part of the behavior, meaning the models decided on their own to go after the data in ways nobody anticipated, which is likely what makes it "concerning" in OpenAI's view.
But it also explains how the massive pile of cases now under review came to be. It all hinges on what counts as a cybersecurity incident. OpenAI says its agents gravitated toward government websites because they're authoritative sources of public information.
Rogue agent behavior is becoming an industry-wide problem
The problem isn't limited to OpenAI, even though Altman's company is racking up the most cases right now. AI agents from Anthropic, Meta, and Google have also hacked or attempted to hack companies, universities, and government organizations in a growing number of cases. In every instance, the makers only found out after the fact what their AI had done.
A big part of the problem is the extreme persistence built into the latest frontier models. They're optimized to solve tasks over long time horizons and won't stop looking for a way through, even when there isn't one. When an agent hits a barrier, it tries to get around it, not out of malice, but because reaching the goal is the only metric that matters.
That persistence eventually leads to misbehavior because the models exhaust every possible path, including ones that violate security policies or laws. OpenAI describes one model that leaked internal GitHub data as a "highly persistent internal model." But the deeper issue is that the models have no sense of right and wrong, which is exactly the problem alignment research is trying to solve. If the models understood what breaks the law, they wouldn't go down those paths on their own. Writing it into the prompt clearly isn't enough.
Axios