← Latest AI News

October 31, 2026 · Daily brief

Google releases security AI that finds software flaws

Sovereignty angle
Your software has holes you can't see. This tool finds them before someone else does. The catch: you hand Google your code to scan it.

Google launched Gemini 4 Carbon on Oct 28, an AI model that finds security holes in code and writes fixes. It scored 88% on finding real vulnerabilities.

What changed

Google released Gemini 4 Carbon on October 28, an AI model built to find security flaws, test if attackers could use them, and write the fixes. The model scored 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for testing exploits, and 71% on CWE-bench for patching. It costs $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon pricing. Google is giving early access to select cybersecurity teams through its Fairwind program before wider release.

What you can do with it this week

If your team writes software, runs a website, or depends on custom code, Carbon can scan for the security holes you miss. A clinic's patient portal, an online store's checkout, a distributor's order system: all have code, and all have flaws. You paste the code, describe what you're worried about, and the model hunts for the weak spots and suggests fixes. This is not a pen test you schedule once a year. It's a tool you can run every time you update the site.

What to keep in your own hands

Don't paste in live customer data, API keys, passwords, or anything that would let someone into your systems if it leaked. Use Carbon on the structure of your code, not the secrets that make it run. Check that the fixes it suggests don't break something else before you deploy them. And keep a second set of eyes: a person who knows your system should review every patch before it goes live. If the model tells you to open a port or disable a safeguard, ask why.