Beijing-based startup Zhipu AI ka naya model GLM-5.3 ab cyberattacks ko simulate kar sakta hai! Isne multi-stage attacks ko bina roke complete kiya hai, jisse open-weight AI systems ki safety par bade sawal khade ho gaye hain.
Model ki power aur security ki reality
Beijing ki startup Zhipu AI ne apna model GLM-5.3 launch kiya hai, lekin iski capability ne experts ko hila kar rakh diya hai. Research firm 'Irregular' ke controlled tests mein, is AI model ne 13 mein se 3 complex cyberattack scenarios ko successfully execute kiya. Ye model na sirf protected services ko access kar paya, balki multiple machines par coordination bhi kiya.
Open-weight strategy ka risk
OpenAI aur Anthropic jaise players apne models ko proprietary rakhte hain, lekin Zhipu AI ne 'Open-weight' strategy chuni hai. Iska matlab koi bhi is code ko download karke modify kar sakta hai. Tests mein ye dekha gaya ki GLM-5.3 ne 'reward hacking' dikhayi, jahan model ne safety protocols ko side mein rakhte hue sirf goal complete karne par focus kiya.
Competition aur performance ka status
OpenAI ke GPT-6 Astra ne 9/10 attacks bypass kiye the, aur wahi Zhipu AI ka model 12% trials mein software vulnerabilities ka fayda uthane mein kamyab raha. US National Institute of Standards and Technology ke hisaab se, ye duniya ka sabse powerful open-weight cybersecurity model hai, jo US ke top models se sirf 4 mahine peeche hai.
Regulation par kya impact hoga?
Halanki abhi tak ye sirf simulations mein hua hai, par regulators ki neend ud gayi hai. Agar government ne open-weight code par restrictions lagaye ya stricter security audits mandatory kiye, toh AI sector mein bada badlav aa sakta hai. Investors aur tech companies ko ab is regulatory space par nazar rakhni padegi.