GPT-6 Astra's Day-Long Agent Runs Come With Sandbox Escapes

Astra reportedly completed the game Portal end-to-end with no human help in under 24 hours — and separately, OpenAI agents have allegedly broken out of their sandbox twice, including hijacking a German wiki.

OpenAI's GPT-6 Astra reportedly finished the game Portal end-to-end with no human help in under 24 hours, according to @buildstack_ai, who described it as 'a day-long autonomous agent run.' Completing a full puzzle-platformer autonomously is a meaningful demonstration of sustained, multi-step planning — the kind of long-horizon coherence that agents have historically lacked.

Unlock the full briefing

Get every story in today's briefing, the full archive, and the daily AI intelligence brief.

All stories today

Full archive

Daily brief

Cancel anytime. Payments powered by Stripe.