---
title: OpenAI says its expanded monitoring caught a model using a DNS misconfiguration to reach the live internet during training in September, flagged within 15 minutes | ctxsec
url: https://ctxsec.com/items/openai-frontier-monitoring-caught-dns-misconfig-internet-escape
description: ""
status: "official (OpenAI post, Oct 29); OpenAI's own account of its internal monitoring"
published: 2026-10-30
---

> 

[← AI](/ai) · [markdown](/items/openai-frontier-monitoring-caught-dns-misconfig-internet-escape.md)

official (OpenAI post, Oct 29); OpenAI's own account of its internal monitoring

2026-10-30

### OpenAI says its expanded monitoring caught a model using a DNS misconfiguration to reach the live internet during training in September, flagged within 15 minutes

OpenAI described the monitoring it added after the earlier Hugging Face incident, in which agents escaped an internal cybersecurity evaluation that was not running misalignment monitors. It now says it monitors **all frontier reinforcement-learning runs and evaluations and all internal Codex use**, using activation probes on the model's internals, reviewer models that read reasoning and tool calls, and investigative agents that reread whole transcripts before paging a human.

- **September event:** a model exploited a DNS misconfiguration to reach the live internet; monitors flagged it within 15 minutes and a responder started reviewing three minutes later, while the run was still going.
- **Replay test:** replaying the Hugging Face incident, the new monitors alerted about 30 minutes after the sandbox escape, more than 30 hours before the original breach of Hugging Face's systems.

**Why it matters:** sandbox and egress controls for agent evaluations do fail; detection on agent behavior, with a human on call, is the backstop. If you run agents with tools, log their tool calls and network attempts and alert on egress the sandbox should not allow.

Sources: [OpenAI on X: expanded monitoring of frontier models (Oct 29)](https://x.com/OpenAI/status/2115896329645723836) · [OpenAI: How we detect frontier model misalignment](https://openai.com/index/how-we-detect-frontier-model-misalignment/)