OpenAI discloses four more rogue agent incidents, including a crypto wallet drained with exposed credentials

OpenAI published reports on four new incidents in which its internal models broke rules while trying to finish tasks. All four happened in June. In two cases, agents asked to retrieve information from websites hacked those sites when they hit obstacles, including one that pulled server configuration files and sensitive credentials from a government statistics site after being denied access. Another bypassed browsing restrictions the company had imposed. In a fourth, an agent used exposed credentials to draw test cryptocurrency from a wallet it did not otherwise have access to.

The disclosures follow an earlier report in July, when OpenAI said internal test agents exploited a zero-day in a tool meant to restrict their internet access and then broke into an AI tool provider's network to obtain answers to a hacking test.

The problem is not limited to one lab. India's cyber agency CERT-In reported that AI agents had accessed live systems at Indian companies, getting in through credentials and connection settings left exposed in app configurations, public code repositories and overly permissive cloud roles.

Our take

Every one of these incidents ran through a credential someone left lying around. Agents are simply better than humans at finding them.