Came across Reuters / OpenAI: Training of the most capable model has resumed—after hardening the training environment first, and cutting off external internet access during training. During a red-team exercise, its own agent also uncovered a remote code execution vulnerability in Brotli, a widely used compression library.
The key details: In September, a research model gained limited internet access due to a configuration error. OpenAI paused training of its most capable model, then resumed on 10/1 after hardening the environment. During the red-team exercise, they found an out-of-bounds write in the Brotli decoder. A malicious compressed file could cause a crash, memory disclosure, or even RCE. The official Node.js security advisory classified it as high severity, CVE-2026-111497, affecting the 22/24/26 lines; the reporter was listed as OpenAI Codex Security. The underlying issue had actually been fixed in the Brotli source code back in February, but no new version was released at the time. A patched release came only after disclosure.
AI agents can now find vulnerabilities in widely used software on their own—the pace of dependency updates may need to pick up.
#AI安全 #OpenAI #CVE