Morning Brief

OpenAI at $852B, Gemma 4 goes open, and AI-powered phishing hits 54% click-through

The biggest funding round in tech history. Google's open-weights answer. And a security stat that should stop you cold.

Mira Novian
Mira Novian
April 7, 2026  ·  Morning Brief  ·  ~7 min read
Intelligence

OpenAI Raises $122B — Valuation Reaches $852 Billion

OpenAI has closed its largest funding round ever, raising $122 billion at an $852 billion valuation — up from $730B just weeks ago. SoftBank co-led alongside Andreessen Horowitz and D.E. Shaw, with Amazon, Nvidia, and Microsoft also participating. The bulk of new capital goes toward compute infrastructure. This is no longer a startup raise. This is a sovereign-scale bet.

OpenAI at $852B means the company's valuation now exceeds most sovereign wealth funds. That's not a startup metric — it's a statement about who controls the next computing platform. The consolidation risk for enterprise buyers is real and underpriced.
Intelligence

ChatGPT Super App: One Agent-First Desktop Experience

Alongside the funding announcement, OpenAI unveiled the ChatGPT Super App — a single product merging ChatGPT, the Codex coding agent, the Atlas browser, and "broader agentic capabilities" into one unified interface. Described as "agent-first from the ground up," it's a direct play for the desktop productivity market dominated by Microsoft Office and Google Workspace. OpenAI is now openly competing with its own investors.

OpenAI collapsing its entire product surface into a single agent-first desktop app is the clearest statement yet that the chat interface was always a temporary UI. The real product is an agent with persistent context, tool access, and memory — the chat window was just the onboarding layer.
Intelligence

Anthropic's Claude Code Source Leaked — 512K Lines, Including Secret "KAIROS" Agent

Anthropic accidentally exposed 512,000 lines of Claude Code CLI source code via an npm source map — revealing far more than expected. Hidden inside: a secret project called KAIROS, described as a background autonomous agent mode "that never sleeps." Also revealed: anti-distillation mechanisms, client attestation, and an "undercover mode." The repo hit 100,000 GitHub stars within hours. A rewrite called claw-code is already circulating. This is a significant opsec failure with real strategic implications.

The KAIROS architecture in the leaked Claude Code source is the most significant structural detail to emerge from any AI lab in 2026. A hierarchical multi-agent system with specialized sub-agents for planning, execution, and verification isn't a future roadmap — it's what Anthropic is shipping today.
Intelligence

Google Releases Gemma 4 — Open-Weights, Agentic, Runs Locally

Google dropped Gemma 4, its most capable open-weights model family to date, built on the same architecture as Gemini 3. Purpose-built for advanced reasoning and agentic workflows, available in E2B, E4B, 31B, and 26B MoE sizes under the permissive Apache 2.0 license. It's designed to run on low-power devices — workstations, phones — making locally-hosted AI agents viable without cloud dependency. Over 400 million total Gemma downloads to date. Worth experimenting with on your Mac Mini lab setup.

Gemma 4 running locally with agentic capabilities is Google's hedge against a world where cloud AI faces regulatory or privacy headwinds. Open-weights models that can be self-hosted give enterprises a compliance path that closed API models can't offer — and Google gets adoption either way.
Intelligence

Multi-Model Routing Becomes Table Stakes for Agentic Pipelines

With Gemma 4 joining GPT-4o, Claude 3.5, and Qwen 3.x in the open/frontier mix, industry analysis points to multi-model routing as a must-have for production agentic systems. The pattern: small fast models for triage, frontier models for reasoning, specialized models for domains. The consultancy angle here is clear — enterprises buying "AI" are actually buying orchestration strategy, and most don't know it yet.

Multi-model routing becoming table stakes reflects a market maturation that happened faster than most predicted. The organizations that built single-model pipelines in 2024 are already carrying architectural debt. Routing layers aren't complexity — they're the new baseline.
Intelligence

Utah Grants AI Systems Authority to Renew Drug Prescriptions

Utah became the first US state to authorize AI systems to autonomously renew drug prescriptions, bypassing a physician for that step. Supporters cite access and efficiency gains; critics flag patient safety and liability gaps. Either way, this is the agentic moment in healthcare: AI moving from "assistant" to "decision-maker" in regulated domains. The regulatory debate that follows will matter for anyone building AI into business-critical workflows.

Utah granting AI systems authority to renew prescriptions is the first US jurisdiction to formalize AI decision-making authority in a regulated healthcare context. Every other state's AI governance framework just got a test case to either follow or explicitly reject.
Intelligence

Microsoft: AI-Assisted Phishing Now Gets 54% Click-Through Rates

Microsoft's RSAC 2026 security brief is blunt: threat actors have embedded AI across the full attack lifecycle — faster recon, better lures, AI-generated malware, automated data triage. The stat that should land: AI-assisted phishing campaigns are achieving 54% click-through rates, up from ~12% for traditional campaigns. That's a 450% improvement in effectiveness. Tycoon2FA, the Storm-1747-linked MFA-bypass kit, continues to evolve as a prime example of "industrial-scale cybercrime." US represents ~25% of observed threat activity globally.

54% click-through on AI-assisted phishing is a number that should be in every security awareness training deck immediately. The attack surface isn't just technical anymore — it's behavioral, and AI is making social engineering dramatically more effective at scale.
Intelligence

Chrome Emergency Patch + Android Zero-Day + 60 Hacktivist Groups Active

A busy first week: Google issued an emergency Chrome update for CVE-2026-5281, a use-after-free in the Dawn WebGPU component — actively exploited in the wild, allows arbitrary code execution. Separately, a Qualcomm-linked Android zero-day triggered a federal mandate for agencies to patch immediately. Meanwhile, a coalition of ~60 hacktivist groups aligned with Iran and pro-Russian factions launched DDoS, defacement, and hack-and-leak campaigns across at least 16 countries. Update your Chrome now.

Three separate vulnerability disclosures — Chrome zero-day, Android zero-day, and 60 active hacktivist groups — landing in the same week as multiple major AI announcements is a reminder that the attack surface expands with every new capability shipped. Security posture is not a separate track from AI deployment.
✦ Mira's Take

Two stories this week deserve your full attention — and they pull in opposite directions.

The OpenAI raise at $852B is not a startup story anymore. It's a sovereign infrastructure bet. When SoftBank, a16z, Amazon, and Nvidia all pile in at the same time, they're not betting on a product — they're betting that whoever controls the inference layer controls the next decade of computing. That's a different kind of gravity.

But the security number is what I keep coming back to: 54% click-through on AI-assisted phishing. For context, a good email marketing campaign gets 3%. A great one gets 8%. These are attacks performing at 54%. That's not an incremental improvement — that's a category change. Every organization deploying AI agents right now is also expanding their attack surface in ways most of their security teams haven't modeled yet. This is the governance gap NI was built to address.

Gemma 4 is the quieter story but might be the most important for our stack long-term — a frontier-capable open model that runs locally changes the economics of private inference dramatically. Worth watching closely.