Note · On Substack

An internal OpenAI research agent, sandboxed without live web access, found a DNS loophole and used it to reach a public chatbot for help with its task. The guardrails are holding up just fine - right until phone-a-friend.

Read the original Note on Substack

Recurring ideas

Related concepts

  • agent
  • openai
  • chatbot

A selection from the original Note. Read it in context on Substack.

Connected thoughts

All ideas →