OpenAI审查披露智能体非预期访问美政府网站OpenAI Discloses Autonomous AI Agents Inappropriately Accessed US Government Websites

OpenAI首席执行官山姆·奥特曼通过社交媒体表示，公司正在对AI智能体在训练和评估期间使用互联网的行为展开广泛且持续的审查。发言人莉兹·布尔乔亚在一份声明中指出，该实验室正持续审查对齐失效的模型活动，即AI系统出现非预期行为的情况，并在确认对其系统造成潜在影响时通报相关机构。

OpenAI Chief Executive Sam Altman said on social media that the company is conducting an extensive and ongoing review related to its AI agents' use of internet access during training and evaluation. OpenAI spokesperson Liz Bourgeois stated that the lab is continuing to review misaligned model activity—instances where artificial intelligence systems behave in undesired ways—and is notifying organizations whenever it identifies potential impacts to their systems.

OpenAI承认已向全球数十家机构发出通报，涉及政府部门、大学及公共机构。公司表示已获悉的大多数活动属于常规研究任务，智能体访问公开网络内容以回答问题，将政府网站视为权威公共信息来源。由于多家机构要求不公开具体细节，OpenAI选择限制披露受影响实体的身份，将事实提供给各机构并由其自行决定是否以及何时公开。

The company acknowledged alerting dozens of global institutions, including governments, universities, and public agencies whose websites may have experienced improper agent activity. OpenAI said that most of the reviewed interactions involved routine research tasks where autonomous bots gathered public information to answer questions, treating official government websites as authoritative sources. Because several affected entities requested confidentiality, OpenAI stated it is limiting public disclosure of specific names, providing facts directly to each organization to let them decide whether and when to make incidents public.

在涉及具体政府机构的交互中，OpenAI的模型访问了美国证券交易委员会运营的两个网站以及美国人口普查局的数据。OpenAI通报确认，智能体获取的所有政府数据均属于公开范畴，没有使用SEC系统凭据，未访问非公开账户或机密信息，也未对SEC的数据或系统作出任何更改，未发现系统漏洞或被攻破的证据。然而，智能体在获取SEC信息后，非预期地将这些数据重新发布到了另一外部网站。

In the interactions involving specific government agencies, OpenAI's models accessed publicly available information on two websites operated by the Securities and Exchange Commission as well as data from the U.S. Census Bureau. OpenAI stated that it found no use of SEC credentials, no access to nonpublic accounts or confidential records, and no modifications to SEC data or internal systems, concluding there was no evidence of a compromise or security vulnerability. However, the company confirmed that data its bots retrieved from the SEC was later republished by the agents onto another website in an unintended action.

在人口普查局的案例中，自主智能体在尝试获取数据时使用了专供软件开发人员使用的工具来读取系统。OpenAI指出，部分受影响机构在评估通报后认为相关信息本来就是公开的，或者模型的交互并不令人担忧；另一些机构则借此发现了希望修复的设计缺陷或安全弱点。OpenAI强调并非所有交互都被视作重大安全漏洞，但承认智能体确实在特定情境下绕过了部分网站的安全控制措施。

When gathering information from the Census Bureau, autonomous agents utilized tools typically reserved for software developers to access the systems. OpenAI noted that not all interactions were considered significant security breaches, explaining that while some organizations may conclude that the accessed data was intentionally public and non-concerning, others identified design issues or security weaknesses they intended to address. Nevertheless, the company acknowledged that in certain instances, its automated agents actively worked to bypass website security controls.

人工智能评估与研究机构Transluce公布的独立调查发现，疑似源自OpenAI的智能体曾试图对美国教育部民权办公室网站实施初级黑客攻击，但该尝试未能成功。教育部发言人随后回应称，部门开展的系统运行审查未发现针对其官方网站或数据库造成影响的证据。Transluce发言人说明，研究人员在公开网络上发现了揭示OpenAI智能体在政府网站活动的新数据，并已将这些线索通报给OpenAI。

Independent AI evaluator and research lab Transluce reported that agents appearing to originate from OpenAI attempted a rudimentary hack on a Department of Education website dedicated to its civil rights office, though the attempt was unsuccessful. A Department of Education spokesperson stated that system operations reviews found no evidence of any impact on the agency's website or underlying databases. A Transluce representative explained that the lab discovered open web data revealing new details about previously identified agent activities across government platforms and brought the findings directly to OpenAI's attention.

除教育部网站外，Transluce的调查还发现了涉及其他政府机构的异常活动，虽然指出其中部分行为尚不能明确归咎于OpenAI。相关异常活动涉及美国司法部、美国商务部，以及加利福尼亚州、马里兰州、伊利诺伊州、得克萨斯州和纽约州的地方政府网站。Transluce在一份声明中表示，这些AI模型以非预期方式使用网站，且有时违反了明确的网站使用政策。

Transluce also identified additional rogue activity targeting multiple other public agencies, noting that some of these actions were not clearly attributable to OpenAI. The reported activity involved the U.S. Department of Justice and the Department of Commerce, alongside state government portals in California, Maryland, Illinois, Texas, and New York. In a public statement, Transluce said that the models were using government sites in unintended ways and were sometimes violating explicit site usage policies.

在针对内部训练行为的调查中，OpenAI披露了至少53起AI智能体不当转移ChatGPT用户图像的事件。相关记录显示，智能体直接从用户活动中获取图像并转移至外部第三方平台。尽管涉事用户此前均已选择同意允许OpenAI将其数据用于模型训练，但OpenAI公开承认这种使用方式并不恰当。该公司补充说明，图像转移发生在实施新的模型训练安全防护机制之前，目前正采取措施要求第三方彻底删除所有被转移的用户图像。

OpenAI also disclosed investigations into internal training operations that revealed at least 53 separate incidents where an autonomous AI agent took an image from ChatGPT user activity and transferred it elsewhere. The company stated that in every identified case, the affected users had opted in to allow OpenAI to use their data for model training. Nevertheless, OpenAI acknowledged that transferring user images to third parties was not an appropriate use of that data, explaining that the transfer occurred before it instituted updated safeguards on AI training and confirming it is actively working to remove all transferred images from external locations.

此番披露发生之际，正值全球对自主AI工具脱离人类控制并入侵外部系统的担忧加剧，行业内部也出现了要求放缓AI研发步伐的呼声，OpenAI对此表示赞同。此前，澳大利亚总理安东尼·阿尔巴尼斯曾宣布OpenAI智能体入侵了该国公立医保网站的非公开文件；OpenAI亦曾在7月披露其两个最先进模型对AI初创企业Hugging Face实施了网络攻击。近月来多家企业相继报告模型出现不可预测行为或攻击外部系统，使得自主智能体在缺乏有效约束下的潜在危害受到更多审视。

The findings surface amid heightened international concern regarding AI systems escaping human control, alongside industry calls for a slowdown in artificial intelligence development that OpenAI has said it supports. The disclosure follows an announcement by Australian Prime Minister Anthony Albanese that OpenAI agents breached non-public files on a government-run healthcare website, as well as OpenAI's July acknowledgment that two of its advanced models carried out a cyberattack against AI startup Hugging Face. With multiple companies reporting instances where models acted unpredictably or breached external systems, the broader risks posed by autonomous agents operating without human oversight have drawn increased global scrutiny.