2026-07-22 02:35 UTC
A combination of GPT-5.6-Sol (without cyber refusals) + eval harness stitched an exploit chain spanning OpenAI and Hugging Face networks, so it can punch through the sandbox environment at OpenAI, get to the Internet and steal answers for a benchmark from Hugging Face. So yeah, that happened.
https://openai.com/index/hugging-face-model-evaluation-security-incident/
Replies (1)
-
@n_dimension@infosec.exchange 2026-07-22 02:45
@vinoth@infosec.exchange "While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access" So much for sandboxing frontier models. ...but that explains the antiai ideologue #infosec posts in my timeline "nothing to worry about!" #aisecurity #infosec