The AI Hacking Era Has Started

On 21 October, PHP pushed out an emergency fix for two bugs that let a stranger run their own code on your web server. No password needed. PHP runs roughly 70% of the websites in the world, including WordPress, most online shops, a lot of government portals, and this website you're reading now.

Yesterday, Polymarket's users watched their money drain out of their accounts for about an hour. This morning, Binance said someone walked off with around US$6.3 billion, the biggest crypto theft ever recorded. Three very different stories. They all point at the same thing.

A 14-year-old bug, found by AI

The security world calls it PHP2Shell (CVE-2026-109618 and CVE-2026-109619). It's a memory bug that sits in every PHP version from 5.4 up to 8.5. Reports say the faulty code had been there for about 14 years. Plenty of very good engineers worked on that code over the years without spotting it.

AI didn't walk past it. In August, a small AI security firm called Calif found the first part using one of OpenAI's models. A few weeks later, Wiz found it on its own with a mix of Google and OpenAI models. OpenAI then pulled together a group with PHP, Cloudflare and US government agencies to fix it quietly before going public.

That part of the story is good news. AI helped the good guys find a serious hole and fix it. Wiz called it "an AI inflection point", and they're right. But think about what it means. If AI can find a bug that humans missed for 14 years, it can find the next one too. And the next person to find one might not call the PHP team first.

The secret didn't stay secret

The plan was to keep the bug quiet until the fix was ready. It leaked anyway. The bug IDs showed up early with the wrong severity score, a demo video went online, and one vendor shipped its fix ahead of everyone else. News broke on the night of 20 October, a day before the official patch.

Amazon says its traps caught hackers trying the bug on 17 October. That's four days before anyone could patch. Once the news was public, Cloudflare's CEO said attack attempts went to tens of thousands within 45 minutes, hundreds of thousands within two hours, and millions by that evening. Some government agencies in the Netherlands and Finland simply pulled their websites offline.

This isn't a one-off. Google's Mandiant team tracks how long it takes from a bug being announced to hackers using it. Look at how fast that number has dropped:

Now compare that with how most companies patch. Somebody notices a news article. They forward it to the vendor. The vendor replies that it's "in the backlog" and they'll look at it during the next maintenance window. That might be next month. The attacker started last week.

Then the other languages started falling

PHP wasn't the end. Once everyone pointed AI tools at their own code, the reports came in fast. In the four days after PHP's fix, 16 serious bugs in major programming languages and tools were announced and patched. Ten of them involved Node.js, including one researchers nicknamed Node2Shell. Python fixed a certificate check that could let a fake server pass as a real one. GitLab patched a bug that scored a full 10 out of 10 for severity. Java had its own warnings.

PHP even found a third way to run code on servers two days after the first fix. One of the Node.js maintainers said they've already fixed a record number of bugs this year and expect it to get much worse from here.

That makes sense. Every big language has millions of lines of old code that nobody has looked at closely in years. AI is very good at reading old code with fresh eyes, and it doesn't get tired. The good guys will find a lot of bugs this year. So will the bad guys. The only question for you is how quickly your systems get the fixes.

Binance: one laptop, six billion dollars

Binance is the biggest crypto exchange in the world. It has a huge security team and more money for protection than most countries. Early this morning it confirmed that around US$6.3 billion was taken. The old record was the US$1.5 billion Bybit theft last year.

So how did they get in? According to Binance's first update, it started with a laptop belonging to one DevOps engineer. From that laptop, the attacker reached a service inside Binance's systems and sent withdrawals that looked real. Binance has since said it found more break-ins beyond that laptop, and analysts say wallets on about ten different blockchains signed for the attacker within half an hour. Blockchain trackers like TRM Labs and Chainalysis believe North Korea's TraderTraitor group is behind it, the same group the FBI blamed for Bybit. Binance hasn't confirmed that yet.

Tether, Circle and others have frozen close to US$3 billion, and Binance says it'll cover user losses from its own funds. The full story will take weeks to come out. But the starting point is already clear, and it's very ordinary. One person's computer. One set of access that could do too much.

Polymarket: the script nobody thinks about

Polymarket got hit a day earlier, on 29 October, and the way in was even more boring. Like most websites, Polymarket loaded a small piece of code from an outside company called DebugBear. Its job was to measure how fast the page loads. Harmless, or so everyone thought.

Attackers took over the script on DebugBear's side. Every Polymarket user who opened the homepage now ran the attacker's code inside their own logged-in session. The script checked if you were signed in, sold your positions, and moved your money out using Polymarket's own buttons. On-chain trackers counted about US$2.7 million taken from around 700 accounts in roughly an hour, before Polymarket pulled the script and paused trading.

Polymarket didn't write the bad code. It didn't even run the server it came from. It just trusted someone else's script, the same way your website probably trusts a chat widget, an analytics tag, a cookie banner, a payment plugin and a font. Each of those can be turned against your customers if the company behind it gets hacked.

What these stories have in common

None of these attacks needed a movie-style hacker typing green text in a dark room. The way in was a forgotten old bug, one laptop with too much access, and a small script from another company. What changed is the speed. AI makes it cheaper to find holes and turn them into working attacks, then try them on millions of websites at once.

We wrote in The Golden Age of Hackers that AI was making it easier to ship unsafe code. This month showed the other side. Even code written carefully by experts, years ago, is now fair game. A well-built system still gets into trouble if it's left alone long enough.

For a Malaysian SME, this is closer than it looks. Most local business systems we see run on PHP. A lot of them are on versions that stopped getting updates years ago. Many were built by a vendor who has moved on, sitting on a server nobody wants to touch. We wrote about that kind of server in The Server Nobody Could Turn Off. In this new era, that server is the first one to go.

AI works for both sides

We're not anti-AI. We use it every day, and it's the reason PHP2Shell got found and fixed at all. The same tools can scan your code, watch your logs and flag weird logins faster than any human team. Defenders have the same weapons as attackers now.

But a tool only helps if someone is using it. The attacker's AI is running right now, testing millions of websites, and it doesn't care that you're a small company. It goes after whatever hasn't been updated.

So "we'll look at security after the next feature" is no longer a reasonable plan. Security used to be something you could push to next year. Now the gap between a bug and an attack is shorter than your vendor's reply time. Plan for that, or the attackers will plan it for you.

Facts in this article come from public reports by PHP, Fortune, Help Net Security, BleepingComputer, Sekoia, Bitquery and others, as of 30 October 2026. The Binance and Polymarket stories are still developing, and the numbers may change.