PHP2Shell: the AI was the fast part, the humans were the bottleneck

0:05Host:MBA Training. Today: PHP2Shell: the AI was the fast part, the humans were the bottleneck. So the French Interior Ministry, the Port of Rotterdam and 4chan all went dark on the same Wednesday.

0:20Expert:Worst group chat ever assembled.

0:23Host:Australia's own cyber security agency pulled its website too. The people whose job is telling everyone else not to panic.

0:31Expert:In fairness, it was the responsible kind of panic. Two flaws in PHP, the language that builds a web page on the server every time someone clicks. It runs about seven in ten websites.

0:45Host:And the headline everywhere is "AI found it." OpenAI more or less took a victory lap.

0:51Expert:That's the headline I want to argue with.

0:54Host:Oh, come on. The AI found a bug that sat there for years. That is the story.

1:00Expert:Hang on. Humans had already spotted it years ago, and it got flagged again in a security audit two years ago.

1:07Host:Wait, what? Spotted and just... left there?

1:11Expert:Parked. The way the researchers at Calif tell it, nobody could prove an attacker could actually use it, so it never got patched.

1:20Host:The "can't prove it, won't fix it" pile.

1:23Expert:Every organisation has one. Security teams, finance, the data team's backlog. A drawer full of things that are probably fine.

1:33Host:"Probably fine" is my favourite risk rating.

1:36Expert:So what changed this summer is the price of proof. Frontier models, meaning the most capable AI systems available, helped researchers turn that old memory bug into remote code execution. That's the polite term for a stranger running their own code on your server without logging in.

1:54Host:Fine, so the AI did the hard part. I still don't see how that weakens the headline.

2:00Expert:Because the AI was the fast part. After that the humans took over, and it got slow and messy.

2:07Host:Define messy.

2:08Expert:Chainguard, a company that sells locked-down software packages, hears on a Friday that something is coming in PHP. That same evening their team finds the same flaw with their own setup and a frontier model. Their chief executive, Dan Lorenc, says the prompt was basically "find bad vulnerabilities in PHP."

2:30Host:Ah yes, the masterclass in prompt engineering.

2:34Expert:Short enough for a sticky note. No role, no context, no examples to copy from.

2:40Host:I'm trying it on our budget tomorrow. "Find bad assumptions in next year's plan."

2:45Expert:You'll get a list longer than the plan.

2:48Host:So finding it got cheap. What broke?

2:51Expert:The embargo. That's the agreed silence among the people who know about a flaw, so the fix ships before attackers hear about it.

3:00Host:The quiet period.

3:01Expert:Right. And it leaked from every direction. The identifiers turned up on forums. Then Calif accidentally published a video of its exploit, the actual working attack, running against several targets. They deleted it, but copies had already spread.

3:19Host:Oh no.

3:20Expert:The PHP Foundation pulled the release forward a day. Then on release morning, Chainguard published patched images, ready-made bundles of the fixed software, more than six hours before the agreed time.

3:33Host:Honestly... I think I'd have done the same.

3:36Expert:Go on, defend him.

3:38Host:Vendors in the group said attacks were already happening. The details were all over the place. And he made the images free for everyone, customers or not. If the house is on fire, you don't wait for the scheduled fire drill.

3:51Expert:Except his fire extinguisher came with a map. Once a fix is public, anyone can compare the old code with the new code and see exactly where the hole was. Reuters quotes experts saying AI tools can now reverse engineer the bug from a public fix with little effort.

4:10Host:So the cheap proof works for attackers too.

4:13Expert:The price dropped on both sides of the table. The Foundation's wording was that Chainguard "unilaterally broke the embargo."

4:21Host:And Lorenc apologised.

4:23Expert:For the extra stress on the maintainers. He defended the decision itself and said the call was his.

4:30Host:Peak executive apology. Sorry you're upset, I'd do it again.

4:35Expert:I'll give him one thing: he owned it publicly, under his own name. In most companies a bad call dies quietly in a committee nobody can name.

4:45Host:But you still think he was wrong.

4:47Expert:I think he cited exploitation reports without naming the vendors. The Foundation says its team found no working public exploit by release time. So it's his reading against theirs.

5:00Host:And I think that's a long morning to sit still while people tell you attacks are live. We're not settling this one.

5:06Expert:We're not.

5:07Host:Can I say how the maintainer side hits me? A small team doing open-heart surgery on the web's engine while the whole internet keeps hitting refresh. That's brutal.

5:18Expert:Yeah. And it wasn't a quick fix. The flaws sit in memory management and error handling, which almost every request goes through.

5:27Host:So who actually loses here? I don't think it's the governments that went offline for a day.

5:32Expert:No, they did the sensible thing. The quiet losers are the servers nobody owns anymore. PHP only patches its supported versions. Censys, a firm that scans the internet, counted more than half of the PHP servers it sees running a version with no official fix.

5:51Host:More than half.

5:52Expert:End-of-life versions, meaning the maintainers stopped fixing them a long time ago. The microsite from some old campaign. The investor page a contractor built and nobody has opened since.

6:05Host:Hey. Marketing microsites have feelings.

6:08Expert:They also have admin passwords from the previous decade.

6:12Host:Fair. My problem with your framing, though: you're saying AI isn't the story, the humans are. Isn't that a bit comforting? "Relax, it's just old-fashioned process failure."

6:23Expert:It's the opposite. If AI were the whole story, you'd buy an AI tool and feel safe. If the bottleneck is human, then every company's pile of parked risks just got repriced, and nobody sells a product for that.

6:38Host:Hmm. Okay, that's uncomfortable in a useful way.

6:42Expert:Picture the finance chief's risk register, the running list of known risks. Some items are rated "unlikely" partly because proving otherwise would take serious time and money. That cost just collapsed, for attackers too.

6:58Host:So "unlikely" sometimes really meant "expensive to check."

7:02Expert:For years. And the old rule made sense: fix what you can prove. It stopped working this summer, and nobody updated the rule.

7:12Host:This is where I'd send people to our lesson "Data crisis management: incidents, communication and resilience." This whole week was a communication crisis wrapped around a technical one.

7:23Expert:And a resilience one. The crisis didn't start on Tuesday with the leak. It started when that ticket got parked.

7:31Host:I want to push on the hype, though. Someone online was already calling it "Black October."

7:36Expert:And the vendors were already writing the webinars.

7:40Host:"Agentic, AI-native, zero-trust PHP shield." Book a demo.

7:46Expert:Meanwhile the real story is a small foundation, a parked ticket and a broken quiet period. Nobody's building a dashboard for that.

7:55Host:To be fair to the AI camp, OpenAI said that as far as it knows, frontier AI found these flaws before attackers did. If that's true, defenders won the first round.

8:06Expert:If. On release day, tech firms were already warning that some hackers were exploiting it.

8:13Host:So finding it first and fixing it first are two separate races.

8:17Expert:And the second one runs at human speed. Committees, release schedules, a contractor who left and took the server password with them.

8:26Host:A professor in Australia called it the biggest cyber security emergency in a decade.

8:32Expert:Maybe. Or just the first one we noticed. If proof is cheap now, every parked ticket in every organisation is a headline waiting to happen.

8:42Host:Which takes me back to my sticky-note prompt.

8:44Expert:"Find bad assumptions in next year's plan."

8:48Host:Upgraded version: "Find the bad assumptions we parked."

8:51Expert:That one will ruin somebody's Monday.

8:55Host:If finding the problem is close to free and fixing it still takes a small team months, who actually gets to decide which parked risk comes out of the drawer first, before somebody outside the building decides for them? Sources: Globalbankingandfinance, iTWire and Axios. Every link is in the show notes. That's the Friday wild card. Our lesson "Data crisis management: incidents, communication and resilience" goes deeper, free at mba-training.com.