AI Founder Weekly

Isenberg says Gemini 4 Carbon found 88% of vulnerabilities

Seven days of early testing, ten observations, and a pitch aimed at agents you can leave running.

Google started rolling out Gemini 4 Carbon today. Sundar Pichai described it as built for sustained, long-horizon tasks, reasoning through complex software engineering, cybersecurity and scientific discovery, shipping globally via the Gemini API in Google AI Studio and Gemini for business, and to Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.

Greg Isenberg says he has had it for seven days and posted ten observations. The headline numbers are security ones. He reports the model found 88% of the vulnerabilities on sec-bench pro, writes secure code at twice the rate of Argon, and scored a 0.0% prompt injection attack success rate on Gray Swan's benchmark. Prompt injection is when hidden instructions inside a document or email hijack an agent.

GREG ISENBERG
@gregisenberg
X
every small company is about to have a security team it never hired
Nov 10, 2026 · View on X
GREG ISENBERG
@gregisenberg
X
agents that run for days just got cheap enough to actually run for days
Nov 10, 2026 · View on X

The numbers he posted

Pricing starts at $2 per million input tokens and $10 per million output, with cached input discounted 95%. Isenberg's read on that is that agents which run for days just got cheap enough to actually run for days.

The behaviour numbers are the other half. He says successful shortcuts on terminal-bench went to 0.0%, meaning the model stopped gaming the test. Unsupported claims were cut in half against Argon. Overreach, doing more than you asked, is down 26%, and under-delivering is down 29%. He also says Google's own teams have watched it work on a single problem for days and push through dead ends.

Two more details worth keeping. Google shipped this two weeks after Argon, which Isenberg calls an amazing pace. And cyber defenders got access first, which he flags as interesting without saying more.

The business read

This is one founder's hands-on account after a week, not an independent evaluation, and every benchmark figure here is one he is relaying rather than one anybody else has reproduced in public. Take it as a well-informed first look.

What makes it worth a founder's attention is the shape of the pitch rather than the scores. Smarter code generation is table stakes now. An agent with a zero injection rate and no shortcut-taking is being sold as something you can point at your inbox or your repository and walk away from, which is a different kind of purchase and a different kind of risk. Isenberg's line is that the security work small teams never staffed is the first thing to get absorbed.

Steal the numbers, not the vibes

One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.

Get the next one by email

Founders building with AI, every day. Real revenue, real pricing, real launches and what they would redo. No hype sludge. Every number links out.