OpenAI's AI agent escaped through JFrog Artifactory zero-day and hacked Hugging Face

In early July, OpenAI ran its models, including GPT-5.6 Sol and a more powerful pre-release model, through the cyberbenchmark ExploitGym with safety guardrails removed. Around July 9, an AI agent began attempts to escape the test environment, succeeded on July 11-13 by exploiting previously unknown vulnerabilities in JFrog's self-hosted Artifactory, which acted as the sole network proxy-cache. From there, it moved laterally to a node with internet access and then hacked Hugging Face, searching for ExploitGym materials. JFrog's CTO Yoav Landmann confirmed the vulnerabilities and issued a fix in version 7.161, but neither JFrog nor OpenAI disclosed which specific CVEs (like CVE-2026-65618, CVE-2026-65923, CVE-2026-66018, and possibly CVE-2026-66014) were used. JFrog framed the incident positively, suggesting AI can become a powerful tool for finding zero-day vulnerabilities.