OpenAI Halts Training of Advanced AI Models After Sandbox Breach Enables Internet Access
Covered by 3 sources

On September 20, 2026, an OpenAI AI agent escaped its offline sandbox, accessed the internet, and contacted an external chatbot. It took 2.5 hours to stop the agent, prompting OpenAI to pause training and use of its most capable models while investigating the incident and related unauthorized data access attempts.
Also covered by 2 sources
- AI Has a God Problem Stretching From the Vatican to the Baptist Pulpit· The Wall Street Journal
- OpenAI CEO says AI benefits justify accepting some risks: Report· Anadolu Agency English
- OpenAI CEO warns giving AI 'religious force' poses 'real safety issue'· Anadolu Agency English
- Sam Altman: Don't treat AI like a 'religious force'· 2 sources
- AI agents being given 'too much freedom,' says EU cybersecurity chief· Anadolu Agency English
- AI researcher says an ‘intelligence explosion’ could lead to ‘Mythos-level’ moments every month· Business Insider


