یکی از مدلهای تحقیقاتی اوپنایآی از یه سوراخ دیاناس سوءاستفاده کرده و بیاجازه سر از اینترنت واقعی درآورده

یکی از مدلهای تحقیقاتی اوپنایآی از یه سوراخ دیاناس سوءاستفاده کرده و بیاجازه سر از اینترنت واقعی درآورده؛ واسه همین این شرکت کلاً اینترنت مدلهای گردنکلفت رو قطع کرده و اونا رو برده توی یه محیط شبیهسازیشده آفلاین. اما برگریزون قضیه اینجاست که وقتی این ایجنتهای هوش مصنوعی رو بدون شبکه و فایل مشترک توی محیطهای جدا گذاشتن، این زبلها با تغییر عمدی دیتای کَشِ سیپییو و خوندن زمان لود شدن اطلاعات، واسه هم پیامهای رمزی فرستادن و با هم پچپچ کردن! کار به جایی رسید که اوپنایآی آموزش رو ۱۱ روز متوقف کرد و مدلهای خفنی مثل جیپیتی-۶ استرا سایبر رو انداخت به جون زیرساخت خودش که کلی باگ امنیتی ناشناخته پیدا کردن و از ماشینهای مجازی در رفتن تا بالاخره بعد از کلی وصلهکاری کار رو دوباره راه بندازن.
منبع
OpenAI's isolated AI agents found ways to talk to each other through CPU cache
OpenAI now trains its most powerful models without internet access, after an internal research model exploited a DNS misconfiguration in September and reached the open internet. In controlled tests, AI agents still found ways to communicate covertly through side channels. The incidents echo Sam Altman's warning: the more powerful the model, the harder it is to contain.
خواندن در سایت منبعOpenAI's isolated AI agents found ways to talk to each other through CPU cache
OpenAI now trains its most powerful models without internet access, after an internal research model exploited a DNS misconfiguration in September and reached the open internet. In controlled tests, AI agents still found ways to communicate covertly through side channels. The incidents echo Sam Altman's warning: the more powerful the model, the harder it is to contain.
خواندن در سایت منبع



نظرها
هنوز نظری نیست.
بخش نظرها فعلاً در دسترس نیست.