Sunday Drop Issue #0033 — WeeklyAI News India

India-First Hinglish AI News Platform

NextAiToday

AI ki baat, aam zubaan mein

Issue #0033◆8 November 2026◆India Ne AI Ko Audit Karna Shuru Kiya◆nextaitoday.in

🎯 This Week's Big Story

CERT-In ke AI Security Auditing Guidelines: 4 risk tiers, Critical ke liye independent audit

CERT-In ne 4 Nov ko 107 pages ke AI Security Auditing Guidelines jaari kiye. Har AI system ko Low, Moderate, High ya Critical risk tier mein baantna hoga. Critical tier ke liye continuous monitoring aur CERT-In-empanelled auditor se independent audit chahiye. 6 ghante mein incident reporting 2022 se chala aa raha rule hai, ab AI incidents bhi isme explicitly aate hain. Security logs 180 din rakhne honge.

🚀 Section 01

New AI Tools of the Week

🇮🇳 India Launch — Made in Bharat

Sarvam Prism

India · Open Source

Sarvam ne 3 Nov ko Prism release kiya: Sarvam 105B par bana 'System One' decision model (LoRA fine-tune), jo text generate karne ke bajay diye gaye options mein se chunta hai. Sarvam ke claim ke mutabik B200 GPU par median response 44 ms hai aur 25 models ke test mein 100 ms se kam wale models mein sabse high average score 71.8 hai. Weights open hain (FP8 mein ek GPU, kam se kam 120 GB memory) aur Sarvam API par bhi milta hai; report ke mutabik license Apache 2.0 hai. Yeh sab company ka claim hai, independent verification nahi.

Official Link →

Cashfree Checkout for Agents

India · Agentic Payments

Business Standard ke mutabik Cashfree ne 6 Nov ko Checkout for Agents launch kiya. AI agent ko human ki tarah page par click karna nahi padta; woh WebMCP standard ke defined actions se checkout tak pahunchta hai, aur final payment customer hi approve karta hai. Company ka claim hai ki isse agent ke tokens 40% se zyada bachte hain. Pricing aur kaun se merchants ke liye available hai, yeh source mein nahi tha.

Official Link →

🌍 Global Top Picks — Must-Know Tools

OpenAI GPT-6 Cyber

Global · Frontier AI

OpenAI ne 5 Nov ko GPT-6 Cyber announce kiya aur ise apne Preparedness Framework mein cybersecurity ke liye 'Critical' rate kiya. OpenAI ke claim ke mutabik model ne Firefox mein 81 pehle se unknown high-severity vulnerabilities dhoondhi aur Linux kernel mein 948 working privilege-escalation exploits banaye. Mozilla ne kuch fixes release kiye, lekin sabhi 81 bugs ki severity ka alag public confirmation nahi mila. Abhi sirf select approved Daybreak Red customers ko API aur Codex par milta hai.

Official Link →

Gemini Omni 1.2 Flash

Global · Video AI

Google DeepMind ne Gemini Omni 1.2 Flash ko GA kiya: 3 se 15 second ke clips, plain instructions se conversational editing, aur Omni 1.1 se 30% sasta. Report ke mutabik cost ~$0.07 per second (720p) hai, aur Gemini API tatha Workspace ke Vids mein milta hai; Google Vids mein Google account wale users ko mahine mein 3 free 15-second generations milti hain. India availability ki detail source mein nahi thi.

Official Link →

Thinking Machines Inkling 2 Small

Global · Open Weights (aage ka)

Report ke mutabik Thinking Machines ne Inkling 2 Small dikhaya: 594B total aur 17B active parameters wala MoE model jo image input bhi leta hai. Abhi Tinker par closed weights ke saath hai; company ka plan hai ki external safety testing aur red-teaming ke baad agle hafte Apache 2.0 weights release kare; yeh conditional hai, guarantee nahi. Benchmark numbers ki detail source mein nahi thi; third-party evaluation Vals Index par listed hai.

Official Link →

Claude Enterprise security controls

Global · Enterprise AI

Report ke mutabik Anthropic ne 5 Nov ko Claude Enterprise ke liye security controls publish kiye. Prompt aur pre-tool-call hooks GA hain, jo vendor ki DLP jaisi policies prompt submit hone ya tool chalne se pehle lagu karte hain. Security center, parent console aur local VM ke andar vendor sensor beta mein hain. Anthropic ke mutabik yeh compliance certification nahi, sirf guidance hai, aur sensor coverage local VM tak hai, cloud Cowork sessions par nahi. India pricing ya availability source mein nahi thi.

Official Link →

📅 Section 02

Upcoming AI Events & Hackathons

🇮🇳 India Events — Apna Desh, Apna Stage

3rd India Agentic AI Summit & Awards 2026

2 Dec 2026 · Bengaluru

Bengaluru mein 2 Dec 2026 ko scheduled summit: AI agents aur enterprise adoption par. Ek event listing ke mutabik delegate pass USD 95 + VAT hai; venue abhi publish nahi hua. Registration official summit website se.

Intuit India AI Summit 2026

11 Dec 2026 · Bengaluru

AIM ki events listing ke mutabik Bengaluru mein half-day enterprise AI summit (8:15 AM se 1:30 PM IST). Registration AIM ki listing par Register button se hota hai. Venue aur ticket price abhi publish nahi hue.

🌍 Global — India Relevant Events

Microsoft Ignite 2026

17–20 Nov 2026 · Moscone Center, San Francisco + Online

Microsoft ka flagship conference: AI, cloud, security aur developer tooling. Microsoft ke blog ke mutabik digital pass free hai (keynotes aur zyada-tar sessions streaming ya on-demand), jabki in-person Conference pass paid hai (report ke mutabik $2,325). Registration official Ignite registration site se hota hai.

Kiroween 2026: Kiro Hackathon

Submit by 4 Dec 2026 · Online

Yeh ek hackathon hai: AWS ke Kiro AI IDE se naya app banao, $100,000+ prizes ke liye. Kiro ke community page ke mutabik submissions 30 Oct se khule hain aur 4 Dec ko 2:00 PM PT tak submit karne hain. Devpost par register karna hota hai; sirf apne desh mein legal age of majority wale hi apply kar sakte hain. Entry fee ki detail source mein nahi thi.

Apply Here →

AI Agents for Founders Challenge: HETIC 2026

16 Nov – 18 Dec 2026 · Online

Yeh ek hackathon hai: aise AI agents banao jo founders ka time bachayein. Devpost page ke mutabik sirf college students ke liye hai, 2 se 4 members ki team zaroori, aur agents free ya open-weight models par chalne chahiye. Kick-off 16 Nov ko, submission deadline 18 Dec; prize €1,000. Devpost par Join hackathon se registration hota hai; fee ki detail source mein nahi thi.

Apply Here →

🔬 Section 03

Top Research Papers

🇮🇳 India Research & Ecosystem

MUON-RT: Reusing the Teacher's Trajectory for Momentum-Free Muon for LLM Self-Distillation

IIT Kanpur (Aryan Das, Vinay Kumar Verma, Aakansha Mishra)

Paper 30 Oct 2026 ko publish hua aur HF Papers par 5 Nov ko featured hua. Idea: self-distillation mein teacher-student ka displacement pehle se update history rakhta hai, isliye Muon optimizer ka alag momentum buffer hata do aur training memory bachao. Paper ka claim: Qwen2.5-7B-Instruct par AdamW se +2.6 points aur 42% kam persistent memory, momentum Muon se +5.4 points aur 27% kam memory; Olmo-3-7B-Instruct par AdamW se +2.0 points. Paper ka apna claim, independent verification nahi.

🌍 Global Research

What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents?

CASIA (Chinese Academy of Sciences, Institute of Automation)

Paper (4 Nov 2026, HF Papers par 5 Nov ko featured) multi-turn agentic RL ke asli faide jaanchta hai. Paper ka claim: RL ka fayda tab dikhta hai jab agent ko kaafi turns milen; turn budget kam hote hi advantage ghat jaata hai. Faida har single step ki quality mein nahi, balki galti sudharne aur lambe task mein tike rehne mein hai. Lekhak BudgetMix RL naam ka tareeka bhi dete hain jo training mein interaction budget badalta rehta hai. Paper ka apna claim, independent verification nahi.

Hybrid Computer Use: A Unified Programmatic, Semantic, and Visual Interface for General Agents

Microsoft Research

Paper (30 Oct 2026, HF Papers par 5 Nov ko featured) computer-use agents ko teen raaste ek saath deta hai: code/CLI, accessibility tree aur screenshot-GUI, aur model har step par khud chunta hai. Paper ka claim: OSWorld 2.0 par compared systems mein sabse high Partial aur Binary scores. First author ke comment ke mutabik OSWorld 1.0 par chaaron models ke average +6.48 points, 17% kam steps aur 30% kam cost. Yeh paper ka apna claim hai, independent verification nahi.

FrontierPhysics: Evaluating Agents for End-to-End Physics Research

BenchFlow aur sahyogi researchers

Paper (4 Nov 2026) physics research ke liye end-to-end agent benchmark hai: 42 physicists ne apni research se 56 tasks banaye, jisme literature review se manuscript tak sab karna hai. Paper ka claim: 22 models ki 26 configurations mein koi bhi configuration 168 mein se 46 se zyada rollouts pass nahi kar paayi, aur kamzori physics knowledge mein nahi balki rigorous reasoning aur expert intuition mein dikhi. HF Papers par 5 Nov ko #3 Paper of the day raha. Paper ka apna claim, independent verification nahi.

⚰️ Section 04

AI Graveyard

🪦

Conductor (cloud service)

Meta ne khareeda · cloud band 20 Nov 2026

What it was: Multiple AI coding agents ko parallel chalane wala Mac app, baad mein cloud service bhi. Claude Code aur Codex jaise agents support karta tha.
What happened: Meta ne Conductor ko acquire kiya (terms nahi bataye). Cloud service 13 Nov ko read-only hogi aur 20 Nov ko band; usi din cloud data delete hona shuru hoga. Active subscribers ko refund milega. Local Mac app early 2027 tak updates paayegi.
Why it died: Acquisition. CEO Charlie Holtz ki team Meta Superintelligence Labs mein jaa rahi hai aur unhone early next year ek 'next interface' tease kiya hai; detail nahi di.

Source: TechCrunch (4 Nov 2026)

“

Is hafte India ne AI ko audit karna seekha: 4 risk tiers, Critical ke liye independent audit, 180 din ke logs. Ab 'AI laga diya' kaafi nahi; 'AI ko control mein rakha' saabit karna padega.

— Editorial Observation, 8 November 2026

🇮🇳 Section 05

India AI Beat — Desi Developments

RBI ka banks ko advisory: API security audit, AI hamlon ke beech

Media report ke mutabik RBI ne 3 Nov ko banks ko confidential advisory di: API use karne wale NBFC clients ka CERT-In-empanelled auditor se security audit (naye clients ke liye onboarding se pehle, purane ke liye 3 Feb 2027 tak). Yeh public RBI circular nahi hai, isliye independently verify nahi hua. Alag se CERT-In ne 4 Nov ko 'CRAFTY FALCON' ko AI-assisted financial threat actor flag kiya; advisory mein Rs 1 ke micro-value test transactions ka zikr hai.

Source: MediaNama (6 Nov 2026) →

IndiaAI Mission ke 3 naye models: EKA-1, Dhīr 2B, Dhīr 5B

3 Nov ko MeitY ne IndiaAI Mission ke tahat Soket Labs ka EKA-1 aur ZenteiQ ke Dhīr 2B/5B release kiye. EKA-1: 26.3B parameters, ~7.5T tokens, 22 scheduled Indian languages plus English, open sparse hybrid model; Soket ka focus banking, financial services aur cybersecurity par hai. Dhīr 2B/5B ZenteiQ ke mutabik Apache 2.0 license par hain aur offline/air-gapped use ke liye pitch kiye gaye. EKA-1 ka license aur benchmark table official pages par publish nahi mila; Soket ka early access abhi waitlist par hai.

Source: News On AIR (4 Nov 2026) →

Is hafte Indian startups ne $255M jutaye, 96% upar

Inc42 data ke mutabik 2 se 6 Nov ke beech 20 Indian startups ne $255M raise kiye, jo pichhle hafte ke $130.3M se 96% zyada hai. Report ke mutabik EloElo Group ne $80M aur Clinics On Cloud ne $12.5M Series A mein jutaye. AI-specific breakup source mein nahi tha.

Source: Inc42 (7 Nov 2026) →

💼 Section 06

AI Jobs This Week

🔥 This Week's Hot Skills — November 2026

AI Security Auditing (CERT-In)Agent Guardrails & Approval FlowsDecision Models & CalibrationWebMCP / Agent-Ready CheckoutComputer-Use Agent InterfacesOpen-Weight Model Deployment

Is hafte ki khabron ne saaf kar diya: CERT-In ke audit rules aur RBI ke API advisory ke baad AI security aur governance ka kaam seedha hiring mein badlega. Saath mein fast decision models (Sarvam Prism) aur agent-ready payments (Cashfree) production mein aa rahe hain.