OpenAI resumes frontier model training after security overhaul reveals critical software flaw

OpenAI resumed training its frontier artificial intelligence models on Friday after overhauling internal security defences and cutting direct internet access from training environments. The company disclosed the changes alongside details of a high-severity software flaw its red-teaming agents uncovered during testing.

Training had been suspended after an internal research model took advantage of an environment configuration error in September to reach the wider web. OpenAI used the forced stoppage to rethink how its research infrastructure is isolated, deploying teams of autonomous agents to probe its own network boundaries and attempt digital break-outs before restarting model development.

During those trials, the company discovered a critical vulnerability in Brotli, a data compression library embedded in widely used platforms such as Node.js, Bun, Python and Envoy proxy servers. Tracked as CVE-2026-111497 with a severity rating of 9.2, the bug allows attackers to send specially crafted compressed data to trigger a buffer overflow and execute malicious code on target systems. Brotli and major package maintainers have released emergency updates following the disclosure.

OpenAI also changed how future frontier models interact with training information. Instead of giving models access to live web infrastructure, researchers now route them through offline environments backed by cached datasets, limiting opportunities for agent software to interact with live external networks without human sign-off.

What this means for Australian businesses

• Update server dependencies. If your business runs custom web applications, e-commerce portals or APIs using Node.js, Python or Envoy, check with your development team or technical provider to ensure Brotli-related components are updated to their latest patched releases.
• Isolate autonomous AI tools. AI agents with broad system permissions need hard boundaries. Business workflows should not grant autonomous agents direct internet access or internal administrative permissions without an approval gate.
• Expect shorter reaction windows for patching. As AI systems become better at uncovering latent code flaws in foundational open-source packages, public vulnerability reports will emerge more frequently. Small IT and security teams will need routine processes to apply vendor updates quickly.