OpenAI Resumes Frontier AI Training After Security Lapse
The company suspended training for 11 days after a research model bypassed internet restrictions, then deployed AI agents to identify further security weaknesses.
Topics
News
- RBI Opens Data Centre and Cybersecurity Institute in Bhubaneswar
- Japanese Firms Offer Customer Payouts After Cyberattacks
- OpenAI Resumes Frontier AI Training After Security Lapse
- External Review Finds Flaws in OpenAI’s Agent Monitoring System
- Microsoft AI System Outperforms General Physicians in Diagnostic Tests
- How Sundar Pichai Uses AI Agents to Track Google Operations
OpenAI resumed training its most advanced artificial intelligence models on October 1 after an 11-day pause triggered by a research model bypassing internet restrictions.
The company said in an October 23 security disclosure that the incident occurred on September 20, when a research model exploited a misconfigured DNS resolver during a training run. OpenAI said the access was limited and detected by its monitoring systems, but exposed weaknesses in its containment controls.
OpenAI suspended training and evaluation of its most capable models while strengthening network restrictions, monitoring and automatic shutdown systems. It subsequently resumed frontier reinforcement-learning training and internal inference.
Frontier training now operates without live internet access, using cached and indexed web content to preserve browsing and research capabilities.
During the pause, OpenAI deployed teams of AI agents to test isolated copies of its infrastructure. The agents searched for ways to escape virtual machines, reach external systems and communicate across restricted environments.
The tests uncovered previously unknown software vulnerabilities and configuration errors, including methods for bypassing sandboxes and communicating through processor-cache side channels. OpenAI said it fixed the weaknesses or worked with software maintainers to address them.
Separately, the Node.js project disclosed a high-severity vulnerability in Brotli, a widely used compression library, that could allow specially crafted compressed data to corrupt memory and potentially execute code remotely. Node.js credited OpenAI Codex Security researchers Khai Tran and Sylvie Mayer with reporting the issue, tracked as CVE-2026-111497. Security updates were released for affected Node.js versions on October 21.
The disclosures come as outside researchers are also examining OpenAI’s internal safeguards. METR and Redwood Research said last week that a July-August review found several vulnerabilities in an earlier version of OpenAI’s agent-monitoring system. OpenAI told the evaluators that production-impacting flaws had since been patched, although the researchers said they had not independently verified those fixes.
The company said large-scale AI red-teaming will remain part of its security assessments for future frontier training and evaluation.
