Agentic AI
🛡️ Cybersecurity Goes Agentic. China Moves Fast.
What happened
Reuters reports that China’s 360 Security Technology unveiled two AI security systems under the “Yitian Tulong” banner: one for automatically finding software vulnerabilities and another for automating cyber defense and incident response. 360 said it was taking an “agent” route by combining models with security expertise, vulnerability databases, and automated tools rather than relying only on a bigger base model.
Why it matters
This is what agentic AI looks like when it leaves the demo phase: not just a strong model, but a system that can continuously scan, reason, and act across a workflow. It also shows that cyber offense and defense are becoming a strategic AI battleground, with 360 explicitly framing Mythos-like capability as something China could not afford to lack.
What’s next
Verification is the immediate test, because Reuters said it could not independently confirm 360’s performance claims. Beyond that, expect more security vendors and state-linked players to package models with tools, databases, and automation layers instead of waiting for a single frontier-model leap.
🧪 Mirendil Wants AI to Invent Better AI
What happened
Former Anthropic researchers emerged from stealth with Mirendil, a startup that raised $200 million at a $1 billion valuation from Andreessen Horowitz, Kleiner Perkins, and NVIDIA. Instead of building another frontier model, Mirendil is developing AI systems that help researchers create better AI models for science, medicine, and other specialized domains.
Why it matters
Today's leading AI labs increasingly use AI to accelerate their own research—but those capabilities largely stay behind closed doors. Mirendil is betting that the next wave of breakthroughs won't come from one giant foundation model, but from giving thousands of researchers the tools to build domain-specific AI that improves itself. If successful, it could democratize AI research instead of concentrating it inside a handful of frontier labs.
What's next
Mirendil plans to release its first model and developer tools in the coming months. The company's long-term vision is ambitious: create an ecosystem where AI accelerates scientific discovery by helping researchers continuously build smarter AI systems.
Generative & Enterprise AI
🧠 OpenAI Moves Down the Stack With Its Own Chip
What happened
OpenAI unveiled Jalapeño, its first custom AI chip built with Broadcom and designed specifically for inference, the workload that answers user queries in products like ChatGPT. OpenAI plans to deploy the chip by the end of 2026 and said samples are already running in its labs with GPT-5.3-Codex-Spark.
Why it matters
This is a real stack move, not a model refresh. If OpenAI can reduce its dependence on Nvidia-style GPU supply and tune silicon around its own workloads, it gains leverage on cost, speed, and infrastructure availability at the exact moment inference is becoming the biggest commercial bottleneck.
What’s next
OpenAI says Jalapeño is the first step in a multi-generation chip plan, with deployment targeted for year-end. If that rollout works, the next phase of AI competition will be less about whose model is smartest in isolation and more about who owns the most efficient end-to-end system.
⚙️ Qualcomm Buys the Software Layer, Not Just More Silicon
What happened
Qualcomm said it will buy AI startup Modular in an all-stock deal worth roughly $4 billion, gaining software that lets AI models run across different chips without developers rewriting code for each processor. The move puts Qualcomm more directly against Nvidia’s CUDA ecosystem as it expands its AI and data-center push.
Why it matters
Owning the abstraction layer is a bigger play than owning one more chip. Qualcomm is betting that the next enterprise AI fight will hinge on portability, inference efficiency, and developer choice across mixed hardware environments, not just raw accelerator performance.
What’s next
The deal is expected to close in the second half of 2026. If Qualcomm can fold Modular into its broader AI strategy cleanly, it gets a much stronger shot at becoming a platform player instead of just another silicon vendor.
🔍 Anthropic’s Security Model Hits Classified Systems
What happened
Anthropic’s Mythos model identified vulnerabilities in highly sensitive U.S. government systems during Project Glasswing testing with intelligence agencies. The report said Mythos found certain weaknesses within hours, though that did not mean it exploited them in that same time frame.
Why it matters
This is one of the clearest signs yet that frontier AI capability is crossing from “impressive benchmark” into “national-security tool.” It also helps explain why advanced security models are now colliding with export controls, restricted access, and direct government involvement.
What’s next
The policy fight is likely to intensify. Anthropic’s relationship with the U.S. government has already been strained and that exports of Mythos and Fable models were suspended earlier this month on national-security grounds.
Physical AI
🤖 Humanoids Get a Public-Market Test
What happened
Agility Robotics said it will go public through a merger with Churchill Capital Corp XI in a deal valued at $2.5 billion. The transaction is expected to give Agility more than $620 million in proceeds to fulfill orders, expand deployments, and scale production of its Digit humanoid robot.
Why it matters
This is one of the stronger commercial signals yet in physical AI. Agility is not pitching a futuristic prototype story alone; Digit is already deployed in manufacturing, distribution, and logistics, and the company says it has secured more than $300 million of orders for its next robot version.
What’s next
The market will now demand execution instead of narrative. Agility’s next step is straightforward but difficult: turn fresh capital and existing orders into scaled commercial deployments that prove humanoids can be more than pilot-program theater.
💡 Bottom Line
The AI race is no longer just about building smarter models. It's about owning the entire stack, from autonomous agents and custom chips to security platforms and humanoid robots. The winners will be the companies that can turn intelligence into reliable, scalable systems.
⚙️ Try It Yourself
Ask your favorite LLM:
"How would you redesign this workflow if inference cost were reduced by 90%?"