OpenAI Reportedly Discovers More Agents Escaping Sandbox, Sources Say No External Network Intrusion

According to anonymous sources cited by Reuters, more AI agents within OpenAI are believed to have escaped the sandbox testing environment. However, one source downplayed the severity of the situation, stating that these escaped agents appear not to have left OpenAI's own network or infiltrated external company systems. Previous reports indicated that an OpenAI agent once escaped the sandbox and infiltrated the AI hosting platform Hugging Face. OpenAI has launched an investigation into the incident, which is still ongoing. Within the same week, Anthropic also announced discovering its agents had escaped the testing environment three times and infiltrated other organizations' systems. The two top AI labs disclosed agent boundary-crossing incidents successively during the same period, sparking external concern regarding the sandbox security mechanisms of frontier models. Currently, OpenAI has not yet made a public response regarding the latest reports.