OpenAI is clearly not playing around this month. On Tuesday, April 14, the company unveiled GPT-5.4-Cyber. This is a souped-up, specialized version of the already massive GPT-5.4 model, built specifically to help cybersecurity professionals protect systems from digital villains.
But here is the catch: you cannot just log into ChatGPT and start playing with it.
This one is strictly invite-only. OpenAI is rolling it out slowly to vetted security vendors and researchers through its Trusted Access for Cyber (TAC) program, which first surfaced in February. Only the users who pass the highest tier of identity verification get the keys to this particular kingdom.
So, what makes this model so special?
GPT-5.4-Cyber has been trained to be "cyber-permissive." In plain English? It will actually answer the tough security questions that regular AI models usually dodge out of "safety caution."
One of its coolest tricks is binary reverse engineering. This lets experts peek inside compiled software to hunt for malware, vulnerabilities and security weaknesses without ever needing the original code.
And the timing of this drop? Not exactly subtle. Anthropic announced Claude Mythos on April 7 as part of "Project Glasswing," which is their own controlled defensive initiative. OpenAI followed up exactly one week later. Coincidence? Almost certainly not.
Oh and, OpenAI actually brought the receipts. They cited some insane progress from their Codex Security product. According to their report, this thing has already helped fix more than 3,000 critical vulnerabilities.
And get this: the glow-up is real. They’ve been testing these models on "capture-the-flag" benchmarks (basically the AI Olympics for hacking). GPT-5 was sitting at a decent 27% back in August 2025. Fast forward to November, and GPT-5.1-Codex-Max absolutely smashed it with a 76% score.
Because of this, OpenAI is now treating every new release like it has the potential to reach "High" levels of cyber-capability. They aren't taking any chances with their Preparedness Framework, and honestly, after seeing those numbers from Anthropic, I can’t blame them.
Also in AI This Week: Google Chrome Got a LOT Smarter
Google quietly dropped a feature that your future self is going to obsess over. On the same day as OpenAI’s big reveal, they announced Skills for Chrome, which in my opinion is a game-changer that lets you save your favorite AI prompts and reuse them across different websites with a single click.
Here is the tea on how it works:
Say you always ask Gemini to find vegan substitutions when you’re looking at recipes. Instead of typing that request every single time, you just save it as a "Skill." You can then trigger it on any webpage using a simple forward slash ( / ) or the plus sign ( + ) button. That is it. One click and Gemini is already doing the heavy lifting.
Early testers are already using Skills for:
Calculating protein macros from random cooking blogs.
Comparing product specs while online shopping.
Summarizing endless documents without leaving the tab.
Google is even launching a Skills Library filled with pre-built options for productivity, budgeting, and shopping. So, even if you’re feeling uninspired, there is a ready-made shortcut waiting for you.
The Fine Print: Skills are rolling out now to Chrome desktop users signed into a Google account. The only kicker? Your browser language needs to be set to English (US) for now.
So yeah, go forth and automate your life. You deserve it!
