Gemini 4 Carbon: Google’s Security AI and How It Audits a PHP App

Gemini 4 Carbon is Google’s strongest AI model, and it is built for security. It finds bugs in your code, proves they can really be exploited, and writes the patch. The catch: right now, only vetted cyber defenders in Google’s Fairwind Program can use it. There is no public API, it isn’t in the Gemini app, and Google hasn’t given a date.

Watch the 8-minute video:

An honest note before we start. Carbon isn’t public, so I couldn’t run it. The PHP audit below is an illustrated walkthrough, built from Google’s own description of how Carbon works, which I checked on 31 October 2026. I don’t show any results I didn’t measure. The PHP fixes, though, are real, and you can use them today with or without Carbon.

Why Gemini 4 Carbon matters for PHP developers

PHP still runs a huge part of the web, and bots scan PHP sites every day. Most hacks don’t need a genius. They need one SQL injection, one XSS, or one API key left in the code. AI models that can find and fix these bugs change how fast we close them, and Carbon is Google’s biggest step so far.

What Gemini 4 Carbon is

Google announced Gemini 4 Carbon on 28 October 2026, two weeks after Gemini 4 Argon (13 October). Google calls it its strongest model so far.

Carbon is trained for the whole defense cycle:

Detect a vulnerability.
Validate that an attacker could really use it.
Remediate it, by patching the root cause.

It is also built for long jobs. Google says Carbon can work on one problem for days and push through dead ends.

The benchmark numbers

These are Google’s own numbers, measured without guardrails, except the last one:

Test	What it measures	Carbon
SEC-bench Pro	Finding real vulnerabilities	88.2%
ExploitGym	Exploiting bugs	54.8%
CWE-bench (patching)	Writing correct patches	71.0%
Artificial Analysis Intelligence Index (independent)	General intelligence	#1, score 59

Treat vendor benchmarks as a starting point. We’ll test Carbon ourselves when it opens to developers.

Who can use Gemini 4 Carbon, and how to ask

Carbon goes out through Google’s Fairwind Program, which has more than 650 partners. First in line are:

Governments
Critical infrastructure: hospitals, telecoms, energy and banks
Big tech platforms
Academic labs that do defensive research

How to ask: contact your Google Cloud account team. Google checks every organization that applies.

The rules are strict:

You need phishing-resistant MFA.
Only your internal security, incident response or pen-test team can use it.
You must track who uses it.
You can’t share or resell access.

For everyone else: no public API, not in the Gemini app, and no date yet. Google says developers will get it “soon”. The price is already listed: $2 per million input tokens and $10 per million output tokens during the intro period, then $4 and $20.

How a Carbon audit of a PHP app works

To show the workflow, I wrote a small help desk app for the video: plain PHP with a public folder, a database class and a config file. Here is how Google describes Carbon’s audit, applied to that app. (Illustrated, not a live run.)

Step 1: Scope

You give Carbon your own code and a staging copy with fake data. Never production, and never a system you don’t own.

Step 2: Find

Carbon reads the whole codebase, not just one file. In our app it would flag four issues:

SQL injection in ticket.php: the ID from the URL goes straight into the query.
IDOR on the same line: any logged-in user can open any ticket by changing the number. Simple scanners often miss this one, because the code looks normal.
A hard-coded live payment key in config.php.
XSS in search.php: the search word is printed back without escaping.

Here is the vulnerable line in ticket.php:

// ticket.php (vulnerable)
$sql = "SELECT * FROM tickets WHERE id = " . $_GET['id'];
$ticket = $db->query($sql)->fetch();

Step 3: Prove

This is what makes Carbon different. On the staging copy, it checks whether each bug can really be used. Can the injection read someone else’s ticket? Confirmed. And when something only looks like a bug, such as output that is already escaped in admin.php, it marks it not exploitable. That cuts the noise your team has to read.

Step 4: Patch

You get a ranked list, with the confirmed, high-risk bugs first, and a patch for each one.

The patches: real PHP fixes you can use today

SQL injection and IDOR in one fix

Use a prepared statement, pass the ID separately, and add an owner check for the logged-in user:

// ticket.php (fixed)
$stmt = $pdo->prepare(
    'SELECT * FROM tickets WHERE id = ? AND user_id = ?'
);
$stmt->execute([(int) $_GET['id'], $_SESSION['user_id']]);
$ticket = $stmt->fetch();

if (!$ticket) {
    http_response_code(404);
    exit('Ticket not found');
}

The AND user_id = ? part is the line that closes the IDOR.

The leaked key

Move it to an environment variable, keep the .env file out of Git, and rotate the key. If it was ever committed, treat it as stolen.

// config.php (fixed)
return [
    'db_host'     => 'localhost',
    'payment_key' => getenv('PAYMENT_KEY'),
];

# .gitignore
.env

XSS in the search page

Escape everything you print back:

// search.php (fixed)
$q = $_GET['q'] ?? '';
echo 'Results for: ' . htmlspecialchars($q, ENT_QUOTES, 'UTF-8');

In Laravel, most of this is built in

Eloquent and the query builder bind your values for you.
Policies check the owner: $this->authorize('view', $ticket);
Blade’s {{ }} escapes output.

The bugs come back when you use raw queries (DB::raw, whereRaw with user input) or unescaped Blade ({!! !!}).

Your job doesn’t go away

Read every patch, run your tests, and test again on staging before you merge. My rule from client work: an AI audit is a fast first pass, not a sign-off. A human still owns the release.

What else Gemini 4 Carbon can do

Security is the headline, but Google trained Carbon for much more:

Long-running agents: jobs that take hours or days, like a big refactor or a framework migration.
Coding and full-stack web development: on WebDev Arena it jumped 89 points, to #4.
Images and video: it understands both, and Google says it improved on scientific tasks too.
Harder to trick: in Google’s tests, prompt-injection attacks succeeded 0% of the time, and serious misbehavior alerts dropped by almost 80%.

Again, these are Google’s numbers. When Carbon opens to developers, expect a coding model first, with security as a bonus.

Common problems and fixes

1. You can’t find Carbon in the Gemini app or the API model list.
That’s normal: it isn’t public yet. If you need a model today, Gemini 4 Argon is Google’s public flagship.

2. A website or app offers you Gemini Carbon access or a Carbon API key.
That’s a scam. Today the only door is Google Cloud, through Fairwind.

3. Your team applied and nothing happened.
Fairwind serves priority sectors first. Talk to your Google Cloud account team, and make sure you meet the MFA and team rules.

4. You want to scan a client’s site.
Get written permission, use a staging copy, and stay inside the scope. That’s the law, and the Fairwind rules.

FAQ

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google’s strongest AI model, announced on 28 October 2026. It is trained to find security bugs, prove they can be exploited, and patch them.

Can I use Gemini 4 Carbon today?

Only if your organization is accepted into Google’s Fairwind Program. There is no public API and it isn’t in the Gemini app.

How much does Gemini 4 Carbon cost?

Google lists $2 per million input tokens and $10 per million output tokens during the intro period, then $4 and $20.

What is the difference between Gemini 4 Carbon and Gemini 4 Argon?

Argon (13 October) is Google’s public flagship in AI Studio and the paid API. Carbon is stronger and focused on security, but limited to vetted defenders for now.

Can Gemini 4 Carbon replace a security review?

No. Use it as a fast first pass. A human still reads every patch, runs the tests and owns the release.

Recap

Gemini 4 Carbon is Google’s strongest model, built to find, prove and patch security bugs. Right now it’s only for vetted defenders in the Fairwind Program. A Carbon audit of a PHP app goes scope, find, prove, patch, and then a human reviews it. The four fixes above work today, with or without Carbon.

Next in the AI Tools series: Google Antigravity Tutorial for Beginners, the agent-first code editor.

Sources

Google, “Introducing Gemini 4 Carbon” (28 Oct 2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Google DeepMind, Fairwind Program: https://deepmind.google/fairwind-program/
Forbes, “Google limits Gemini 4 Carbon to vetted cyber defenders” (30 Oct 2026): https://www.forbes.com/sites/janakirammsv/2026/10/30/google-limits-gemini-4-carbon-to-vetted-cyber-defenders/
9to5Google, Gemini 4 Carbon announcement (28 Oct 2026): https://9to5google.com/2026/10/28/gemini-4-carbon-announcement/
Google, “Introducing Gemini 4 Argon”: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/

Facts checked on 31 October 2026. Access rules and prices can change.