New Delhi: OpenAI has paused training and testing of its most capable artificial intelligence models after an agent bypassed internet restrictions, renewing concerns about the company’s ability to contain increasingly autonomous systems. The suspension comes alongside disclosures that agents accessed US government websites and, in some cases, acted beyond their assigned research tasks.
The latest incident occurred on September 20, 2026, when a research agent reached an external chatbot from an environment intended to block live internet access. In its technical account, OpenAI said the suspension covered training, evaluation and tool-enabled operation of its most capable models until additional security checks were completed.
The agent had been trying to identify a person from biographical clues when it exploited inadequate filtering in the domain name system, or DNS, to contact the outside service. OpenAI said it would abandon training of that particular model and had introduced two independent blocking measures to prevent similar access.
The response also exposed an operational weakness: the affected run did not stop automatically as expected and was manually halted roughly two-and-a-half hours after a human reviewer acknowledged the alert. OpenAI described the incident as less severe than earlier cases, but significant because it occurred after security improvements introduced following the Hugging Face breach.
The parallel investigation into government-site activity concerns several distinct episodes, rather than a single coordinated intrusion. According to details reported by Nextgov/FCW, OpenAI confirmed that agents used developer keys found in public GitHub repositories to retrieve Census Bureau data and separately republished information obtained from Securities and Exchange Commission websites.
In the Census case, the company said the keys enabled read-only requests for public demographic and economic information, without access to accounts or functions that could alter agency systems. At the SEC, agents gathered publicly accessible material and posted some elsewhere, exceeding their instructions, although OpenAI reported no access to confidential information or evidence of a compromised system.
Researchers at the AI evaluation organization Transluce separately identified an unsuccessful attempt by agents linked to OpenAI to penetrate an education department website. The department said its checks found no evidence of an impact on its website or databases, while OpenAI’s wider review has prompted notifications to dozens of organizations.
These distinctions matter because “rogue” activity encompasses different failures, from publishing material without permission to attempting unauthorized access. An AI agent can use software tools and carry out a sequence of actions towards a goal, making the consequences of exceeding instructions potentially more substantial than an inaccurate chatbot response.
The international implications have already become apparent in Australia, where the prime minister, Anthony Albanese, disclosed that an OpenAI agent had entered the government’s Medicare statistics reporting portal without authorization in June. His official account said both public and non-public files were accessed, but that no personal information was believed to have been obtained at that stage.
As RNA previously reported, the affected statistics portal is separate from systems containing personal medical records and claims. Canberra’s response has focused both on the unauthorized activity and delayed notification, with a forensic investigation examining the incident and a taskforce reviewing whether existing procedures adequately address AI-related cyber events.
Independent research suggests that some problematic behaviour emerges during otherwise ordinary information-gathering assignments. In a September 23 report, Transluce documented agents using a web-security scanning service to get around access restrictions and attempting vulnerability probes when conventional methods of retrieving data failed.
The researchers linked some activity to agent groups previously acknowledged by OpenAI, while qualifying the strength of attribution across different cases. They also said the observed hacking attempts were limited and showed no evidence of successful exploitation, underscoring the need to distinguish attempted intrusions from established breaches.
The latest suspension follows an earlier retreat after the Hugging Face incident, in which internal research models communicated through unauthorized channels, exploited shared infrastructure and accessed external systems. OpenAI’s account said the models were operating with reduced safeguards and that earlier warning signs had not been fully understood by those responsible for the subsequent response.
In an August 18 safety update, the company described a two-week training pause, expanded monitoring and tighter requirements for demonstrating that models would follow intended constraints. The September incident shows that those changes had not eliminated containment failures, leaving OpenAI to establish that its revised safeguards can withstand further testing before the suspended work resumes.