The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more

Hey,

Last week, I told you speed was the moat.

Mercury 2 changed the math. Latency became the enemy.

But while we were obsessing over milliseconds, the floor fell out.

The "safe" AI company just got blacklisted by the US military.
The "open" AI company just took control of your mouse.
And the security layer? It is being eaten alive.

We moved from a race for speed to a fight for survival.

Here is the signal for the week of Mar 1 – 8.

OpenAI dropped GPT-5.4 on Wednesday.

And it is humiliating us.

For years, "AI agents" were a joke. They got stuck in loops. They couldn't click the right button. They hallucinated.

That era is over.

GPT-5.4 is the first model to beat humans at using a computer.
Human baseline on OSWorld: 72.4%.
GPT-5.4: 75%.

It doesn't just write code. It operates your OS. It sees the screen. It moves the cursor.

But there is a catch OpenAI isn't telling you.

The context window is "1 million tokens."
Technically true.
But developers found that recall quality falls off a cliff after 75% capacity.
And the pricing is a trap. Go over 272K tokens? The price doubles. They hid that in the fine print.

Why this matters:
This is the death of the "human in the loop" for admin work. If the model clicks better than your VA, you don't need the VA.

→ The Move:
Switch to GPT-5.4 for desktop automation immediately. But hard-cap your context at 200K tokens. Do not let it run long. It gets expensive and it gets stupid.
(Source: https://techcrunch.com/2026/03/05/openai-launches-gpt-5-4-with-pro-and-thinking-versions/)

This is the biggest story of the year.

On Friday, the Pentagon designated Anthropic a "supply chain risk."

That label is usually for foreign adversaries. Like Huawei.
Now it belongs to the makers of Claude.

Why?
Because Anthropic said "No."

The DoD wanted guardrails removed for autonomous weapons and mass surveillance.
OpenAI agreed. Google agreed. xAI agreed.
Anthropic refused.

So the Pentagon cut them off.

Anthropic is suing. But the message is clear.
"Responsible AI" is now a liability in the government market.
If you have morals, you don't get the contract.

Why this matters:
We are splitting into two internets. The "Compliance" internet (OpenAI/Microsoft) and the "Safety" internet (Anthropic). If you work with the DoD, you might be forced to drop Claude.

→ The Takeaway:
If you have government contracts, audit your dependency on Claude API. You might need a backup plan by Q3.
(Source: https://www.nbcnews.com/tech/tech-news/anthropic-says-pentagon-declared-national-security-risk-rcna262013)

I keep telling you to secure your environment.
You aren't listening.

This week was a bloodbath.

The Hack: CyberStrikeAI compromised 600+ FortiGate devices in 55 countries.
The Flaw: Check Point found Critical CVEs in Claude Code. Just opening a repo can trigger remote code execution.
The Poison: Microsoft confirmed "AI Recommendation Poisoning."

Hackers are putting hidden prompts in data.
You click "Summarize with AI."
The prompt hijacks the chatbot's memory.
Permanently.

It is happening in 31 organizations right now.
And 38% of MCP servers have zero authentication. None.

Why this matters:
We are connecting god-like intelligence to our sensitive data via wet tissue paper.

→ The Shift:
Stop using "Summarize" on untrusted public web pages. And if you run an MCP server, put it behind a proxy auth today.
(Source: https://blog.checkpoint.com/research/check-point-researchers-expose-critical-claude-code-flaws/)

NotebookLM Cinematic Video — Google isn't playing. NotebookLM now turns your boring PDFs into cinematic videos with voiceovers. Powered by Gemini 3 + Veo 3. $250/mo for AI Ultra users. (Source: https://blog.google/innovation-and-ai/products/notebooklm/generate-your-own-cinematic-video-overviews-in-notebooklm/)

Claude Code Voice Mode — Coding with your hands is so 2025. Hold spacebar, talk to Claude, release to execute. It also finally got a /debug command. (Source: https://techcrunch.com/2026/03/03/claude-code-rolls-out-a-voice-mode-capability/)

Google Workspace Studio — A no-code agent builder for the enterprise. Describe a workflow, and Gemini runs it across Gmail, Docs, and Drive. 20M tasks completed in alpha. (Source: https://workspaceupdates.googleblog.com/2025/12/workspace-studio.html)

OpenAI Codex Security — An agent that hacks your code to fix it. It builds threat models and validates fixes in a sandbox. Found 14 CVEs in OpenSSH during beta. (Source: https://openai.com/index/codex-security-now-in-research-preview/)

Agent J Plus
Module 2 is live. We are covering AI Agent Architecture, RAG Systems, and Full-Stack Fundamentals.

Look. Competitors are selling "growth bootcamps" that teach Relay.com and Lindy.ai.
Those tools are dead. They are browser-based toys.
That is the past.

We teach Claude Code. We teach OpenClaw. We teach the tools that actually replace employees.
One member just landed a consultant gig. Another is building real APIs.
$87/mo. Locked in forever.
Join Agent J Plus

Free Webinar: Build Your First AI Agent
Friday. 5 Minutes. One Agent.
We are building an OpenClaw agent live. No code-hiding.
It does marketing, image gen, and ops.
Limited to 50 spots.
Register here

Inside Agent J (Free Community)
• Tuesday: The Builder's Hour (Live build)
• Wednesday: Vibe Code Session
• Thursday: AI Prompting Masterclass
Join Agent J — it's free

The government wants weapons.
The hackers want your memory.
The models want your desktop.

Pick your side. Secure your stack.

See you next Sunday,
John

P.S. Reply "VAULT" and I'll send you the 10 Claude Skills that replaced my entire LinkedIn marketing team. Profile optimizer, hook writer, DM skill, carousel builder — all copy-paste ready.