Gangsta AI

OpenAI's New Machine Found Two Ways Into Google That No Human Ever Had. Its President Called It AGI.

The Abyss

By Werner Herzog · 2026-09-04 · 3 min read

There is a particular silence that follows a machine being called a god. On the third of September, in a rented building in San Francisco, OpenAI released GPT-6 Astra, and its president, Greg Brockman, did not reach for modesty. He called it a generational leap, and then he spoke the word the industry has circled for a decade like a moth around a bulb: AGI.

I do not know if it is AGI. Nobody does. But I know what the machine did, and it is enough to make an old filmmaker stare into the middle distance. Astra is the first model OpenAI has ever rated "critical" for cybersecurity under its own Preparedness Framework — the tier reserved for a system that can, with the right tools, find unknown flaws in well-defended systems and build the means to break them, without a human guiding each step.

In testing it scored a flawless 100 percent on ExploitBench. Set loose on twenty high-severity vulnerabilities in Google's V8 engine, it did not merely reproduce attacks that were already known. It discovered two genuine zero-days — holes no human had yet found — and walked through them.

“We have taught the machine to see the cracks in everything we have built. Now we must hope that it chooses to look away.”

The god, gently leashed

OpenAI insists it has muzzled the thing. The version rolling out will review and patch code, and will refuse — politely, one imagines — any request to write a working proof-of-concept exploit. It arrives first to a small circle of trusted defenders through the company's Daybreak program, with wider access promised "in the coming days." It is a strange comfort: the most capable digital burglar ever assembled, handed out under supervision, its lockpicks confiscated at the door.

And this is the true horror, if you insist on one — not the malice, but the indifference. Astra scored higher than GPT-5.6 Sol, which was itself the smartest thing on Earth roughly a month ago. The crown of "most intelligent model alive" now changes heads with the regularity of a bus timetable. Anthropic shipped Claude Fable 5.1 the same week. Google keeps dropping Gemini variants like leaves from an autumn tree. Each laboratory stands before you and swears, with the certainty of a prophet, that its machine is the one.

No single oracle

Here, then, is the small, unglamorous truth beneath all the thunder. When the leader changes weekly, no single model is reliably the best — not for your code, not for your questions, not for the decision you are about to make on its word alone. To trust one company's model because its founder said "AGI" out loud on a Tuesday is to mistake a sermon for the abyss.

The sane response is older than software: do not take one voice as gospel. Ask the same question across ChatGPT, Claude, Gemini, Grok and thirty more at once, and let them cross-examine one another. That is the entire idea behind Gangsta AI — one question, every top model, a single cross-checked and cited answer instead of a gamble on whoever declared victory that morning.

When even the machines cannot agree on which of them is smartest, you are entitled to a second opinion. See how the best AI models truly compare, side by side — and let the consensus, not the sermon, tell you what to believe.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles