OpenAI's Next AI Is a Master Hacker, And It's Almost Here

OpenAI is reportedly putting the finishing touches on a new AI model called Astra, and the thing the company keeps highlighting is not writing essays or answering questions — it's cybersecurity. According to the report, Astra has reached what OpenAI internally calls the "Critical Threat" level. In plain English, that means the AI is considered capable enough to find weaknesses in computer systems and use them to break in, largely on its own, without a human walking it through every step.

Why should you care? Because the software Astra is good at attacking is the software you use every day — banking apps, hospital records, email, the websites you shop on. The upside is real: security teams are badly outnumbered, and an AI that scans for flaws in seconds instead of weeks could patch holes before anyone exploits them. The downside is just as real. The same ability, pointed the wrong way, could help scammers write more convincing fake emails, break into accounts faster, or build ransomware that spreads on its own.

This isn't the first AI to help with hacking. Security researchers have used AI for years to hunt bugs and review code. What makes this different is the jump from "assistant" to something that can plan and carry out a multi-step attack with little human help. That's why major AI labs now publish safety rules that trigger extra testing and safeguards when a model crosses specific danger thresholds — and why the word "Critical" in that rating is a big deal, not marketing.

The catch: all of this rests on a report, not a public announcement. OpenAI hasn't released Astra, published the test results, or explained exactly what "Critical Threat" means in practice — and outside experts haven't been able to check the claim. Ratings like this come from companies' own internal safety frameworks, and reasonable people disagree on where the lines should sit. In the meantime, the boring advice still works best: turn on two-factor authentication, install your updates, and slow down before clicking anything urgent.