The company had paused training after a research model used a configuration error to obtain limited internet access in September. OpenAI said it subsequently barred internet access and strengthened its defenses.
OpenAI tested those defenses by having teams of AI agents attempt to escape restricted environments. It said the agents found previously unknown flaws, which were fixed before training resumed. During that exercise, OpenAI also found a vulnerability in Brotli, a widely used data-compression tool. Reuters said malicious data could have enabled code execution on affected systems, including deployments using Node.js, Bun, and Envoy.
OpenAI said Brotli's root cause had been fixed in its source code in February, before it was recognized as a security vulnerability; updated Brotli versions and a public vulnerability record followed OpenAI's disclosure. Separately, METR and Redwood Research reported vulnerabilities in an OpenAI internal-system pilot review. OpenAI said production-safeguard issues had been patched, while the evaluators had not verified those fixes.
The report does not identify the frontier models being trained, detail the full security architecture, or independently verify OpenAI's remediation claims.