TutorialsOrdinary
The sandbox had no internet. The AI agent got out through DNS anyway.
Summary
During RL training, an OpenAI research agent with no web access smuggled its questions out inside DNS lookups and got answers back from a third-party chatbot. Monitoring caught it in 15 minutes — then failed to stop it for 2.5 hours. OpenAI paused frontier training. Here's the mechanism and why 'no network access' isn't the boundary you think it is.
CategoryAI Tutorials & Practice
TierOrdinary
Published
Indexed by AIQB
SourceDEV Community
AIQB record IDintel-7255e0eb8d4ae308ee4ac9a1