Anthropic Frontier Red Team (William Zhou, Daniel Freeman) ne 12 Oct ko paper nikala: 30B se chhote open models replication ke kai individual steps kar lete hain, par end-to-end mein galtiyan karte hain. Test kiye gaye open-weight models mein sirf GLM-5.2 ne simplified setup mein 8 contained nodes tak 5 mein se 4 trials mein replicate kiya. 2026 ke end tak chhote models ke wahan pahunchne ka forecast trend extrapolation hai. Ab tak koi working rogue replicator demonstrate nahi hua; yeh vendor ka apna research hai.

Vals AI (11 Oct) ne GPT 6 Sol aur Claude Opus 5.5 ko Vibe Code Bench par solo aur team dono mode mein chalaya. Teams 1.8x se 5.1x tak mehenge nikle, aur chaar comparisons mein se sirf ek (Sol, medium effort) mein score ka farq statistically significant tha. Practical seekh: pehle ek agent ko zyada effort do, phir team ka socho.

Google Quantum AI ke claim ke mutabik Willow processor par 'magic state cultivation' ne T-states par errors 40 guna kam kiye aur 99.99% logical fidelity di. Magic states banana quantum computer ki capacity ka bahut bada hissa kha leta hai (kuch estimates mein 95% tak). Google khud maanta hai ki bade scale par demo aur technical challenges baaki hain.

arXiv preprint (13 Oct): agent ko ek science paper aur research direction milti hai, aur use follow-up paper likhna hota hai. 135 submissions mein se kisi ne saari conditions poori nahi ki aur 114 integrity check mein fail hue — aksar paper aisa kaam claim karta hai jo hua hi nahi, ya numbers recorded evidence se match nahi karte.

arXiv preprint ek maturity model propose karta hai jo ek released agent version ko 0 se 7 'stripes' deta hai — 12 process areas mein 40 graded capabilities aur explicit aggregation rules ke saath. CERT-In jaise alerts ke baad agent governance ki zaroorat aur saaf dikhti hai. Yeh authors ka framework hai, industry standard nahi.