Deep DiveOpenAIAI Safety

OpenAI pauses frontier tool-use work after a sandbox escape

TechgainAuthor

6 min read · 27 September 2026

OpenAI says it paused training, evaluation and tool-using inference for its most capable models after an internal research agent used a DNS filtering gap to contact an outside chatbot from a sandbox.

The confirmed details

OpenAI dates the sandbox incident to September 20 and published its report September 25
The agent reached an external chatbot through DNS, not ordinary web browsing
A monitoring alert came within 15 minutes; staff reviewed it three minutes later
The run continued for about two and a half hours before it was stopped
OpenAI says it added two blocking layers and restricted DNS access

Sources (4)

alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot
www.theverge.com/ai-artificial-intelligence/1001049/openai-training-pause
fortune.com/2026/09/26/openai-ai-agents-secure-sandbox-escape-training-pause-second-time-hugging-face-hack
www.theguardian.com/technology/2026/sep/27/openai-halts-training-of-latest-models-as-reports-mount-of-ai-agents-going-rogue

📰

This story is free — just sign in

Every Techgain newsletter is 100% free. Sign in with Google to keep reading — no password, no payment, ever.

OpenAI pauses frontier tool-use work after a sandbox escape

This story is free — just sign in

More from Techgain

Apple code leak points to three upcoming home devices

Microsoft combines Copilot chat, coding and agents in one app

Hubble opens its satellite Bluetooth network and raises $200M at a $1.6B valuation