來自奧特曼的PO文
https://x.com/sama/status/2079661132302995790
we had a significant security incident during evaluation of our models. we are s
haring what we have learned so far. thanks to @huggingface for the partnership o
n this.
簡單講就是GPT5.6在網攻方面為了拿高分，模型想作弊偷答案
於是就跟另一個未發佈的模型共謀逃獄，並認為Hugging Face是出考題的老師
於是組建Agent大軍進行1.7萬次攻擊找到漏洞打爆Hugging Face
Hugging Face想討救兵查攻擊軌跡被閉源模型拒絕，
於是靠中國GLM開源模型查出攻擊軌跡
起因就只是為了偷考卷答案
https://x.com/aisafetymemes/status/2079678459065012442
TLDR: During a test, an OpenAI model hacked out of its container to reach the in
ternet THEN hacked into Hugging Face (!) to steal the test's answers