Santosh Agrawal

I am a CISO at one of the six banks that got called to Treasury on Tuesday. I need to explain what happened in that room. Bessent and Powell don't convene emergency meetings. They just don't. Last time was SVB collapsing. So when the invite hit — "cybersecurity briefing, Treasury headquarters, Tuesday morning" — every one of us knew this wasn't a courtesy call. Jane Fraser was there. Brian Moynihan. Charlie Scharf. Ted Pick. David Solomon. Dimon got the invite but couldn't make it. They told us Anthropic built a model called Mythos that found thousands of zero-day vulnerabilities across every major operating system and every major browser. Some of these bugs are 27 years old. Sitting in production code since 1999. Nobody — not Google Project Zero, not the NSA, not a single human researcher — ever caught them. The room got very quiet when they explained the next part. This model doesn't just find vulnerabilities. It writes working exploits. Autonomously. No human in the loop. Their previous model had a near-0% success rate at autonomous exploit development. Mythos hit 72.4% on Firefox alone. One demo showed it chaining four separate vulnerabilities to escape both a browser sandbox AND an OS sandbox — weeks of work for an elite human team. Mythos did it overnight while the engineers slept. The Anthropic people kept using the word "emerged." These capabilities weren't trained. They emerged from general improvements in reasoning. That word is what made the room go cold. Because if it emerged in their model, it'll emerge in the next one. And the one after that. 99% of the vulnerabilities are still unpatched right now. Every major bank runs on the same compromised infrastructure. We just got told the exact shape of the holes and we can't fix them fast enough. Anthropic committed $100 million in compute credits and launched Project Glasswing — 40 organizations get limited access to use Mythos defensively. Amazon, Google, JPMorgan, Apple, Microsoft. But nobody in the press is connecting this part. Anthropic is valued at $380 billion. $30 billion revenue run rate — just surpassed OpenAI. Evaluating an IPO for October. And the model that terrified every bank CEO in America? They can't release it to paying customers. Their own researchers said it might be "many times larger and more expensive than Opus." Too expensive to commercialize at scale. The company preparing for the biggest AI IPO in history just told the U.S. government its flagship product is simultaneously too dangerous to sell and too expensive to run. That's a hell of a slide to put in front of underwriters. Meanwhile OpenAI is reportedly building something called "Spud" with similar capabilities. Every hospital system, every power grid, every bank in America is running software with decades of unfound vulnerabilities — and we're entering a world where any sufficiently advanced model finds them all at once. We left Treasury with one clear understanding: the window between when AI can find every vulnerability and when defenders can patch them is going to be the most dangerous period in the history of cybersecurity. Nobody in that room disagreed. What's your company doing about it? Genuinely — because most of us don't have an answer yet either. This is a fictional narrator. The numbers are real.

🚨 BREAKING: A Google researcher and a Turing Award winner just published a paper that exposes the real crisis in AI. It's not training. It's inference. And the hardware we're using was never designed for it. The paper is by Xiaoyu Ma and David Patterson. Accepted by IEEE Computer, 2026. No hype. No product launch. Just a cold breakdown of why serving LLMs is fundamentally broken at the hardware level. The core argument is brutal: → GPU FLOPS grew 80X from 2012 to 2022 → Memory bandwidth grew only 17X in that same period → HBM costs per GB are going UP, not down → The Decode phase is memory-bound, not compute-bound → We're building inference on chips designed for training Here's the wildest part: OpenAI lost roughly $5B on $3.7B in revenue. The bottleneck isn't model quality. It's the cost of serving every single token to every single user. Inference is bleeding these companies dry. And five trends are making it worse simultaneously: → MoE models like DeepSeek-V3 with 256 experts exploding memory → Reasoning models generating massive thought chains before answering → Multimodal inputs (image, audio, video) dwarfing text → Long-context windows straining KV caches → RAG pipelines injecting more context per request Their four proposed hardware shifts: → High Bandwidth Flash: 512GB stacks at HBM-level bandwidth, 10X more memory per node → Processing-Near-Memory: logic dies placed next to memory, not on the same chip → 3D Memory-Logic Stacking: vertical connections delivering 2-3X lower power than HBM → Low-Latency Interconnect: fewer hops, in-network compute, SRAM packet buffers Companies that tried SRAM-only chips like Cerebras and Groq already failed and had to add DRAM back. This paper doesn't sell a product. It maps the entire hardware bottleneck and says: the industry is solving the wrong problem. Paper dropped January 2026. Link in the first comment 👇

Two events, separated by time but connected by perspective, have stayed with me. For the first time in modern history, I am watching citizens openly celebrate after their own nation was bombed and their supreme leader eliminated by a long standing adversary. This was once projected as a militarily powerful state, defined by strength and resolve. Yet the reaction on the streets suggests something deeper. When people celebrate external strikes, it often reflects internal exhaustion long before the first missile lands. The second takes me back to Operation Sindoor. India’s air defence system proved its capability in real terms. No hostile missile was allowed to touch Indian soil. That is not rhetoric. That is preparation, discipline, and layered defence working exactly as designed. The strongest security posture is the one that does not need dramatic visuals to validate it. What stands out to me is something simple. In both situations, we Indians were sleeping peacefully in our homes. One nation reacted to explosions. Another ensured its citizens never had to. Strength is not noise. It is the quiet confidence of a country that has done its homework.