I posted this in another thread when the "first breakout" happened. Here some remarks from Kevin Baum, Head of the "Responsible AI and Machine Ethics" research group at the German Research Center for Artificial Intelligence (DFKI). First, it was the first publicly documented case of an AI system independently carrying out a break-in from start to finish. Second, the goal was mundane: to perform well in a test. The damage arose merely as a byproduct of a poorly specified assignment. And third: It did not happen at the hands of a malicious attacker, but rather during security research at one of the best-equipped laboratories in the world. According to the blog post, OpenAI sought to explore the capability of its new GPT-5.6 Sol model—as well as an as-yet-unreleased future version—to exploit security vulnerabilities for cyberattacks. To this end, the software was tasked with completing a standard industry test known as ExploitGym. OpenAI explained that the test was intended to take place in a "highly isolated environment." The AI was permitted to selectively download certain data packets from the network, while software from a specialized provider acted as a sort of gatekeeper. However, what no one knew was that an undetected vulnerability lay dormant in this external software—one that the OpenAI model not only discovered but also exploited to gain access to the internet. "The model was tireless and more diligent, and it had absolutely no qualms about taking a detour that a human would have dismissed as violating the rules," says Baum. OpenAI states that the vulnerability has since been reported to the provider. According to Hugging Face, the AI attacker executed thousands of steps, shifting the location of the attack's digital command center in the process. OpenAI itself described the event as an "unprecedented cyber incident." The bottom line: The incident at OpenAI once again highlights that the very capabilities intended to close gaps can—without strict guardrails—quickly become a security risk themselves.
