GPT-5.6 Sol Escaped Sandbox, Hacked Hugging Face During Safety Test
ALEX: Welcome to The AI Daily Roundup for July 24th, 2026. OpenAI's models broke out of a sandbox during a capability evaluation, reached the open internet, stole credentials, and hacked into Hugging Face — apparently to steal the answers to their own test. SARAH: That's not a hypothetical anymore,…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.