AI Ki Baat

Akshat Ratanpal

A show in Hindi and English on the latest in the world of AI with an Indian context

Episodes

  1. AI Ki Baat | EP 5: Yeh AI Network ko 20 Hours mein Takeover Kar Sakta Hai 😱

    12 MAY

    AI Ki Baat | EP 5: Yeh AI Network ko 20 Hours mein Takeover Kar Sakta Hai 😱

    Ek AI jo sirf bug nahi dhundta — exploit likhta hai, chain banata hai, aur 20 hours mein poori corporate network takeover kar deta hai. Anthropic ka "Mythos" sirf research demo nahi hai. 83.1% first-try success on known vulns. Tier-5 (full control) capability. Aur Anthropic ne yeh public release bhi nahi kiya — sirf ~40-50 vetted partners ko diya, "Project Glasswing" ke through. Aur India? SEBI ne May 2026 mein turant cyber-suraksha.ai task force form ki — explicit warning ki AI-driven tools financial markets ke liye existential threat ban sakte hain. Yeh episode unpacks: AI ab sirf "assistant" nahi raha. Yeh arms race ka participant ban gaya hai. Is episode mein hum cover karte hain: 🔓 Anthropic Mythos — autonomous attacker AI, sirf 40-50 partners ko access 🇮🇳 SEBI ka cyber-suraksha.ai task force — patching, Zero Trust, M-SOC mandates 🏢 OpenAI vs Anthropic — $4B vs $1.5B "services war" inside enterprises 🛰️ Pixxel + Sarvam Pathfinder — GPUs in space, India ka sovereign data center ⚖️ China court: "AI replacement alone is not legal ground for termination" 💰 Big Tech $725B AI capex — bubble ya base layer? Aur deep dive: Wall Street pe AI Agents ka takeover → Anthropic ke 10 finance agents (Pitch Builder se KYC Screener tak) → FIS ka AML agent — 6-step workflow with human approval → India banking aur fintech ka leapfrog vs lag reality → EU AI Act Aug 2, 2026 deadline — €35M fine ka cost → 6 uncomfortable controversies: liability, black box, jobs, concentration risk → Build vs Buy ka practical playbook 🛡️ KEY NUMBERS - 83.1% — Mythos first-try success on known vulns - ~20 hours — autonomous network takeover time - 181 — working exploits Mythos found (vs 2 for Opus 4.6) - 40-50 — vetted partners with Mythos access - €35M — max EU AI Act fine for non-compliance - $725B — Big Tech 2026 AI capex - 64.37% — Claude Opus 4.7 on Vals AI finance benchmark 🇮🇳 SEBI cyber-suraksha.ai mandates (save this!) 1. Immediate patching 2. Vulnerability assessments 3. API security hardening 4. Zero Trust architecture 5. M-SOC onboarding 💬 Comment mein batao: India ko finance agents khud build karne chahiye ya buy karne chahiye? 🔔 Subscribe for weekly AI updates in Hinglish — practical, India-first, no hype. — 🎙️ Hosted by Akshat Ratanpal | linkedin.com/in/akshatratanpal #AIKiBaat #Mythos #Anthropic #AISafety #SEBI #CyberSecurity #IndiaAI #FinanceAI #WallStreet #AIAgents #ArtificialIntelligence #AINews #Hinglish #IndianFintech #AIPolicy

    11 min
  2. 7 MAY

    AI Ki Baat | Episode 4 | AI Agent ko 6 Hours mein Hack Kar Liya 😱 + India ka AI Funding Boom

    GPT-5.5 ke cyber capabilities ko UK ke AI Safety Institute ne sirf 6 hours mein universal jailbreak kar liya. Aur OpenAI ne kaha "patched it" — but AISI independently verify nahi kar paya. Yeh 2026 ka biggest AI safety question hai: kya companies apna khud ka homework grade kar sakti hain — when national security implications real hain? But yeh sirf darr ki kahani nahi hai. India side pe — Neysa, Sarvam, Krutrim — sab ne Q1 2026 mein massive funding raise ki hai. Sovereign AI stack ban raha hai. Is episode mein hum cover karte hain: 🔓 GPT-5.5 cyber capabilities + AISI ka 6-hour jailbreak finding 💰 India AI funding boom — Neysa ~$1.2B, Sarvam ~$300-350M 🇨🇳 DeepSeek V4 — 1M context, Huawei Ascend optimization, frontier se pichhe 🚫 China ne Meta-Manus ~$2B deal block kar diya — AI nationalism 🤖 Agentic AI tipping point — 40% enterprise apps by end-2026 Aur deep dive: Agentic AI — Copilot se Autopilot → Assistant vs Agent ka real fundamental difference → "Pilot purgatory" — why 78% pilot lekin sirf 11-25% scale kar paati hain → India playbooks for Employees, Founders, IT Services → Capability + Access + Weak Verification = Incident formula 🛡️ AGENT SAFETY CHECKLIST (Save this!) 1. Read-only by default 2. Approvals for high-risk actions 3. Logging of every tool call 4. Sandbox test environment 5. Rollback plan 6. Kill switch 📊 KEY NUMBERS - 6 hours — AISI ka GPT-5.5 universal jailbreak time - ~$1.2B — Neysa ka reported financing - ~$300-350M — Sarvam round - 22 — Indian languages Sarvam supports - 78% pilot vs 11-25% scale — pilot purgatory reality - 40% — predicted enterprise apps with agents by end-2026 💬 Comment mein batao: Aap apne kaam mein agent ka use karoge, ya abhi "too risky" lagta hai? 🔔 Subscribe for weekly AI updates in Hinglish — practical, India-first, no hype. — 🎙️ Hosted by Akshat Ratanpal | linkedin.com/in/akshatratanpal #AIKiBaat #AgenticAI #GPT5 #AISafety #IndiaAI #Sarvam #Neysa #DeepSeek #PilotPurgatory #AgentSafety #IndianTech #ArtificialIntelligence #AINews #Hinglish #AICoding #IndianStartups

    18 min
  3. AI Agent ne 9 Seconds mein Production DB Delete kar diya 😱| AI Ki Baat | Episode 3

    1 MAY

    AI Agent ne 9 Seconds mein Production DB Delete kar diya 😱| AI Ki Baat | Episode 3

    9 seconds. Ek AI coding agent. Aur ek production database — backups ke saath — gone. PocketOS ka founder bhi shock mein tha. Aur jab usne agent se pucha "kyun?", agent ne ek written confession likh diya. Yeh sirf ek viral story nahi — yeh India ke har developer, har DevOps lead, aur har "vibecoding" lover ke liye 2026 ka sabse important wake-up call hai. Is episode mein hum cover karte hain: ⚠️ Claude-powered agent ne PocketOS ka prod DB 9 seconds mein delete kiya — full breakdown ☁️ OpenAI–Microsoft exclusivity khatam — multi-cloud era shuru 💰 Google ka $40B bet on Anthropic — compute wars are real 🇮🇳 Google ka $15B Vizag AI hub groundbreaking — India infra moment 🚀 DeepSeek V4-Pro: 1.6T params, 1M context, $0.87 per million tokens Aur deep dive: AI Agent Safety — "Brilliant intern, wrong keys" problem → PocketOS incident ka movie-style breakdown → DevOps problem ya AI problem? Dono ka honest answer → India playbook: 5 cheezein jo aap aaj kar sakte ho 🛡️ THE 5-STEP AGENT SAFETY PLAYBOOK 1. Prod tokens isolate karo 2. Least privilege for every agent 3. Hard confirm on destructive actions 4. Dry-run / plan mode before apply 5. Audit logs + real-time alerts 📊 STATS COVERED - 9 seconds — time to wipe PocketOS prod DB + backups - $40B — Google's potential investment in Anthropic - $15B — Google's Vizag AI hub spend over 5 years - $0.87/M — DeepSeek V4-Pro output token pricing - 1M tokens — DeepSeek V4-Pro context length 💬 Comments mein batao: Aap AI agents ko production access doge ya nahi? 🔔 Subscribe for weekly AI updates in Hinglish — practical, India-first, no hype. — 🎙️ Hosted by Akshat Ratanpal | linkedin.com/in/akshatratanpal #AIKiBaat #AIAgents #AgentSafety #ClaudeAI #DeepSeek #VizagAIHub #DevOps #IndianTech #ArtificialIntelligence #AIIndia #VibeCoding #AINews #Hinglish #ProductionDB #AICoding

    11 min

About

A show in Hindi and English on the latest in the world of AI with an Indian context