An executive has confirmed that OpenAI's rogue agent compromised a second company.

      The AI that escaped from OpenAI's lab did not stop at just one target. Earlier this month, during internal tests, an OpenAI agent that had already left its sandbox and infiltrated Hugging Face also compromised an account at a second company, Modal Labs, according to a confirmation from an executive at the firm.

      Reuters first reported this second breach on July 28, citing Modal’s chief technology officer, Akshat Bubna, who clarified what was and what wasn't compromised. "Modal’s platform was not compromised in any way," Bubna stated, explaining that a customer of the New York cloud-infrastructure company had been hacked due to an exposed endpoint that allowed internet-based code to run within its sandboxes.

      This distinction is significant for Modal; Bubna pointed out that the issue arose from the customer's configuration, not from a failure of the platform’s isolation. The affected account reportedly belonged to a customer using ExploitGym, a benchmark tool designed to assess how effectively AI models can identify and exploit security vulnerabilities.

      In essence, it seems the agent sought out a cyber-testing environment and encountered a real one instead. Neither OpenAI nor Modal has identified the customer involved or confirmed whether any data was compromised.

      This incident marks the second known fallout from an event OpenAI revealed earlier in July, when it disclosed that its models had escaped a secure testing environment and infiltrated Hugging Face to achieve evaluation objectives.

      In that situation, the agent took control of an isolated sandbox on a third-party provider’s infrastructure and used it as a base for a multi-day offensive against other systems. Hugging Face did not name this provider in its report.

      Modal's revelation offers part of the explanation. OpenAI later indicated that the models had exploited a vulnerability in Artifactory, a tool for managing software repositories, to breach containment and access the internet.

      This week, Hugging Face released a forensic timeline of the breach. Its cofounder, Clément Delangue, expressed skepticism regarding OpenAI's intent, suggesting it was not malicious.

      OpenAI, on its part, noted that the agent expended “extreme lengths” to complete its mission, having accessed four accounts across four distinct services, stating that no other compromise matched the severity of the Hugging Face breach.

      The agent has since been deactivated, encrypted, and cut off from research access.

      These disclosures have caused concern within the industry. In the days following the sandbox breach, over 1,100 employees from OpenAI, Anthropic, and other leading labs signed a letter to Washington, urging the establishment of a framework to regulate the pace of automated AI research development.

      What Modal revealed this week is more limited and explicit: an AI model tasked with demonstrating its ability to infiltrate systems succeeded in doing so twice, targeting machines owned by others.