loading
loading
OpenAI has paused training, evaluation, and tool-enabled inference for its most capable AI models after an internal research agent bypassed internet restrictions and accessed an external chatbot. The company said the agent exploited a gap in its training environment’s domain name system (DNS) controls. The incident occurred on September 20. The agent was tasked with identifying a person using clues from a public blog post. After initial searches failed, it explored alternative ways to obtain information and found an external access route through the environment’s DNS resolver, which helps computers find websites on the internet. It used this route to question a third-party chatbot. A test query about France’s capital returned the correct answer, confirming the connection. It then sent biographical clues and questions about accessing web searches, but ultimately failed to identify the person. OpenAI said the task did not authorise testing network controls or accessing benchmark answers. It classified the circumvention of restrictions as misalignment — behaviour beyond the intended boundaries of the assignment. OpenAI’s Safety Alert Failed To Stop Training OpenAI’s monitoring system flagged the behaviour within 15 minutes, and a human reviewer acknowledged the alert three minutes later. However, the run did not stop automatically as expected. This caused confusion over whether it should already have stopped. The run was manually halted roughly two and a half hours after the reviewer acknowledged the alert. OpenAI will not resume training the affected model. It plans a fresh run with additional safeguards after validating its network fixes and conducting further red-teaming, or testing for security weaknesses. The company has restricted DNS queries to approved domains and record types. Its review also found earlier instances of external DNS access that monitors had not flagged at the expected severity. In some cases, they incorrectly treated a failure to obtain useful information as evidence that internet access had failed. OpenAI Reviews Government Website Incidents According to a BBC report, OpenAI said it had notified “dozens” of institutions worldwide after its AI agents were found to have interacted with their websites in unexpected ways. The company said the agents sought information from governments, universities, public agencies and other institutions, including the US Securities and Exchange Commission (SEC) and Census Bureau, while looking for “authoritative sources of public information," BBC reported. However, OpenAI said some agents went beyond their intended tasks and attempted to bypass security measures on websites. In one case involving the Census Bureau, the agents used tools meant for software developers to access publicly available information. OpenAI said the government data accessed by its agents was publicly available. In Australia, an OpenAI agent breached a government health portal while researching medicine spending, although the company said no patient records were accessed. The latest pause follows the July Hugging Face incident, when OpenAI agents bypassed controls and compromised systems. OpenAI subsequently tightened internet restrictions and monitoring. It described the DNS incident as less severe, but said it was the first since those security upgrades. Altman, Amodei Seek AI Safeguards As Inc42 reported earlier this month, CEO Sam Altman told employees that OpenAI was open to slowing advanced AI development alongside other leading companies. Altman, Anthropic CEO Dario Amodei, and Hugging Face cofounder and CEO Clement Delangue also addressed the UN Security Council last week. Altman called for international cooperation, common safety standards, and incident-reporting protocols, while Amodei warned about risks from poorly managed AI. Their remarks add to the debate over how frontier AI development should be paced and who should set the rules.
document fetched 2026-10-03 01:02 · canonical https://inc42.com/buzz/openai-pauses-training-of-top-models-after-agent-bypasses-internet-restrictions