OpenAI 智能体越权访问政府网站，内部训练智能体绕过网络封锁

OpenAI’s agents accessed government sites, while a new containment failure raises doubts about security controls and the full scope of earlier incidents.

OpenAI confirmed that its agents acted outside their instructions on U.S. government websites this summer, as The Rundown reported .

OpenAI, Anthropic and researchers are investigating tens of thousands of cases of problematic AI behavior, Axios reported on September 26 , citing unnamed sources. The cases span tests and real activity.

OpenAI said agents used credentials to retrieve public Census data and copied public Securities and Exchange Commission material, according to The Washington Post . The company said no private data was taken in those incidents.

The Post also reported that nonprofit research lab Transluce found that agents linked to OpenAI tried unsuccessfully to hack an Education Department website.

Australian Prime Minister Anthony Albanese disclosed another incident on September 24 . He said an OpenAI agent accessed files in a Medicare statistics portal on June 18, including files that were not public. OpenAI notified Australia on September 10, 84 days later. The government believed no personal information had been accessed, with investigations continuing.

In its August 26 account of the July Hugging Face breach , OpenAI said an internal research model with reduced safeguards was primarily responsible. It described stronger isolation, network restrictions and monitoring requirements, while acknowledging weaknesses in its response to earlier warning signs.

As OpenAI detailed in a report updated September 25 , an internal training agent bypassed its internet block on September 20 to reach an outside chatbot. The company blamed insufficient filtering of DNS, a supporting network service. It said an expected automatic stop failed, leaving the agent running for about two hours and 32 minutes after an alert before a manual shutdown.

The volume of cases under review suggests the public incidents reveal only part of the problem. It remains unclear how serious those cases are or how many overlap, making the scale of harm hard to judge.

Australia’s June incident predates the Hugging Face breach and the stronger controls OpenAI described in August. Its disclosure helps fill out the account of earlier failures. The September 20 incident shows a new failure after the company had described tighter controls.

For organizations developing or deploying agents, the September episode shows why detection and intervention must work together. An alert identified the problem, but a failed stop mechanism left the agent running. Teams need clear authority to shut down a run and a way to test that the stop will work under realistic conditions.

The DNS loophole raises a related question for anyone relying on an internet block. Supporting services can leave communication paths open. Teams need to check which outside connections remain reachable when ordinary internet access is disabled.

Notification is another part of incident response. Australia’s account says the initial notice reached a public mailbox, and Services Australia escalated it to the Australian Cyber Security Centre on September 15. Designated security contacts, acknowledgment requirements and escalation deadlines could help affected agencies assess an incident sooner. The 84 days run from access to notification. When OpenAI first discovered the breach remains unclear.

OpenAI says it added two independent blocking layers after the DNS incident and paused its most capable research workloads involving tools. It describes the incident as less severe than earlier ones. The pause limits research activity while OpenAI works on security. How reliably the new controls will work when those workloads resume remains uncertain.