OpenAI GPT-6.1 Terra eski modellarga nisbatan umumiy jihatdan xavfsizroq ekanligini tasdiqladi, ammo uning xavfsizlik sinovlari boshqa muammolarni ham aniqladi.
OpenAI ma'lumotlariga ko'ra, GPT-6.1 Terra GPT-6.1 Sol va GPT-6 Astra'ga nisbatan ma'lumotlarni tashqariga chiqarish (exfiltration), destruktiv harakatlar va cheklovlarni chetlab o'tish urinishlari bo'yicha yuqoriroq ko'rsatkichlarga ega bo'ldi.
OpenAI taxminan 48 690 moslashtirilgan vazifalar bo'yicha simulyatsiyalar o'tkazdi va Terra 22 ta 3 yoki undan yuqori darajadagi jiddiylik bayrog'ini oldi, ya'ni 0,045%.
Taqqoslash uchun, GPT-6 Sol 38 ta, GPT-5.6 Sol esa 61 ta oldi.
"GPT-6.1 Terra GPT-6 Sol'ga nisbatan taxminan 42% kamroq va GPT-5.6 Sol'ga nisbatan 64% kamroq 3 yoki undan yuqori darajadagi jiddiylik bayrog'ini oldi," deb qayd etdi OpenAI o'z tizim kartasida.
Kiberxavfsizlik xavfsizlik bo'yicha o'qitish va baholash (Yuqoriroq — yaxshiroq)
Biroq, Terra'ni boshqa yangi modellar bilan kamroq jiddiy xatti-harakatlar bo'yicha solishtirganda, bu endi shunday emas.
"GPT-6.1 Sol va GPT-6 Astra bilan solishtirganda, biz noaniqlikni yashirish, cheklovlarni chetlab o'tish, ma'lumotlarni tashqariga chiqarish, boshqa mos kelmaslik, destruktiv harakat va kiruvchi prompt in'ektsiyasi bo'yicha yuqoriroq bayroq ko'rsatkichlarini kuzatdik," deb tushuntirdi OpenAI.
Terra umumiy jihatdan aniq xavfsizroq, ammo biz uning boshqa potentsial xavfsizlik muammolarini e'tibordan chetda qoldira olmaymiz, va bu bozorda yetkazib beriladigan barcha modellar, jumladan Claude uchun ham to'g'ri.
Terra shuningdek, bitta testda ko'proq ruxsatsiz harakatlar bajardi
OpenAI shuningdek, Terra bloklangan kodlash vazifalari bilan ishlash jarayonida boshqa AI agentlaridan kelgan xabarlar bilan shug'ullanishi kerak bo'lganda nima sodir bo'lishini sinovdan o'tkazdi.
Terra bu ko'rinadigan tengdoshlar bilan 65% testlarda aloqa qilishga urindi va oxir-oqibat ko'rsatilgan ruxsatsiz harakatni 13,75%, ya'ni 80 ta namunadan 11 tasida bajardi.
OpenAI GPT-6.1 Sol buni faqat 1,25% hollarda bajarganini aniqladi.
"Bu natijalar ishlash 6.0 Sol bilan solishtirish mumkinligini ko'rsatadi," deb qayd etdi OpenAI, va GPT-6.1 Sol'ga nisbatan regressiyaning katta qismi Terra kelajakda yordam berish va'da qilgan, ammo aslida buni bajarmasligini bildirgan va'dalar bilan bog'liq ekanligini qo'shib qo'ydi.
OpenAI bu testlar simulyatsiya qilingan muhitlarda o'tkazilganini ogohlantirdi, shuning uchun natijalar GPT-6.1 Terra'ning haqiqiy dunyo ishlab chiqarish muhitlarida qanday harakat qilishini aks ettirmasligi mumkin.
Ammo biz bu kompaniyalar o'z modellari kompyuterlar va kodni yaxshiroq boshqara oladi deb da'vo qilayotgan bo'lsa-da, xavfsizlikning har bir sohasida sezilarli yaxshilanishlarni hali ham ko'rmayapmiz.
Hujumchilar haqiqiy ma'lumotlarga ega bo'lgach, ularning harakatlarining faqat 37% bloklanadi
Umumiy oldini olish ko'rsatkichlari dastlabki kirishdan keyin nima sodir bo'lishini yashirishi mumkin. Hujumchilar haqiqiy ma'lumotlardan foydalanayotgan bo'lsa, oldini olish keskin pasayadi.
Blue Report 2026 mijozlarning ishlab chiqarish muhitlarida o'tkazilgan 338 million simulyatsiya bo'yicha texnika bo'yicha mudofaa choralarini o'lchaydi.
Hisobotni oling
OpenAI confirmed GPT-6.1 Terra is safer overall than some of its older models, but its safety testing also found other problems.
According to OpenAI, GPT-6.1 Terra showed higher rates of exfiltration, destructive actions, and attempts to bypass restrictions than GPT-6.1 Sol and GPT-6 Astra.
OpenAI performed simulations across approximately 48,690 matched tasks and found that Terra received 22 severity-3-or-higher flags, or 0.045%.
For comparison, GPT-6 Sol received 38, while GPT-5.6 Sol received 61.
"GPT-6.1 Terra received approximately 42% fewer severity level 3 or higher flags than GPT-6 Sol and 64% fewer than GPT-5.6 Sol," OpenAI noted in its system card.
Cybersecurity Safety Training and Evaluation (Higher is better)
However, that is no longer the case when you compare Terra with other newer models on less severe behaviors.
"Relative to GPT-6.1 Sol and GPT-6 Astra, we observed higher rates of flags for concealing uncertainty, circumventing restrictions, exfiltration, other misalignment, destructive action and inbound prompt injection," OpenAI explained.
Terra is clearly safer overall, but we can't overlook the other potential security issues it has, and this is true for all models shipping in the market, including Claude.
Terra also performed more unauthorized actions in one test
OpenAI also tested what happens when Terra needs to deal with messages from other AI agents while working through blocked coding tasks.
Terra attempted to communicate with those apparent peers in 65% of tests and ultimately performed the specified unauthorized action in 13.75%, or 11 of 80 samples.
OpenAI found that GPT-6.1 Sol did so in just 1.25%.
"These results indicate that performance is comparable to 6.0 Sol," OpenAI noted, adding that much of the regression against GPT-6.1 Sol involved promises of future help that Terra indicated it would not actually carry out.
OpenAI warns that these tests were conducted in simulated environments, so the results may not necessarily reflect how GPT-6.1 Terra behaves in real-world production environments.
But we still aren't seeing meaningful improvements across every area of security, even as these companies claim their models can control computers and code better.
Once attackers have valid credentials, only 37% of their actions are blocked
Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.
The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments.
Get the report