PHP2Shell Flaws Found by AI, and Ramaswamy Ad Backlash

Sam There's a security story today about a pair of flaws in PHP, the programming language that runs a huge slice of the web, and the twist is that AI models were what dug them up. Governments pulled websites offline over it. So what exactly is PHP2Shell, and why did it set off alarms this widely?

Maya PHP2Shell is the nickname for two critical memory-safety flaws in the core of PHP, the server-side scripting language. That means it runs on the web server and builds the page before it reaches you. Wiz, CNN, Politico and TechCrunch all described it as powering roughly 70 percent of websites. The PHP Foundation released emergency patches on 21 October 2026. Each bug scores 9.2 on CVSS v4, a severity scale, and the flaws affect every PHP version from the last decade.

Sam Every version from the last decade sounds like an enormous number of servers. I'd guess the scary part is that someone can break in from outside, but I don't know what they can actually do once they're in.

Maya An attacker who has never logged in can run their own code on your server. They can also read process memory, which may hold passwords, tokens or keys, or simply crash the service. Some accounts say the affected code paths go back to PHP 5.4 in 2012. The fixed releases are 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Older unsupported branches can't just take a patch; they have to move to a current release.

Sam Now the AI part. Did somebody ask a chatbot to find bugs in PHP and it spat these out? And I'm wondering how long the flaws had been sitting there before anyone noticed.

Maya It was more involved than that. Wiz Research says it first found one flaw on 19 August, using its Wiz Atlas agent together with Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol. The next day it confirmed unauthenticated remote code execution against Drupal, and against WordPress the day after. It reported to the PHP security team on 24 August. Wiz said the underlying problems had sat in the code for more than a decade. Separately, the security firm Calif found a critical flaw using GPT-5.6 Sol under OpenAI's Patch the Planet programme.

Sam So a single tool did the whole job. That would mean the human researchers were barely needed, which seems like a stretch to me.

Maya Wiz doesn't say that. Proving the bugs could be exploited took a chain of steps "at the edge of what human researchers could practically achieve." With the models, one researcher finished in a week what would previously have taken a world-class team months. Wiz also warned the same models make it easier for attackers to adapt an exploit across PHP versions.

Sam That cuts both ways, then. If the models can speed up the defenders, they can speed up the attackers too. Did the attackers actually get there first here?

Maya The planned disclosure slipped. The PHP Foundation said details leaked before the intended release through a mixture of mistaken and intentional breaches, and it named Chainguard as having unilaterally broken the embargo, the agreed quiet period before going public. OpenAI wrote that, to its knowledge, frontier AI found the vulnerabilities before attackers did. Amazon said it saw a probe against its decoy systems on 17 October, four days before disclosure.

Sam Four days early is odd, given the patches weren't out yet. And once the news broke on the 21st, I imagine the attack traffic jumped, but I'd like to know by how much.

Maya Cloudflare turned on default web-application-firewall protections, filters that block known attack patterns, at 03:02 UTC that morning. Its chief executive, Matthew Prince, wrote on X on 22 October that the network saw thousands of exploit attempts in the ten hours before the announcement. Within 45 minutes afterward there were tens of thousands, within two hours hundreds of thousands, and by his post, millions. Cloudflare also said attackers were mutating their payloads to slip past its rules.

Sam Two camps are reading this episode differently, I gather. One says it proves AI helps the good guys, the other says it hands attackers a head start. Who is on which side, and what are they pointing to as evidence?

Maya OpenAI and Wiz present it as proof that frontier models can find, and help fix, flaws humans left untouched for years. They call it a defender's window. Critics and some practitioners accept the same fact and read it the other way: once details leaked, the same class of tools makes it practical for attackers to adapt exploits quickly.

Sam What would settle it? I'd think a confirmed break-in at a real, named website would be the proof either way.

Maya That is the open point. As of 23 October, government statements and vendor telemetry, meaning the traffic data security companies collect on their own networks, had not fully converged on whether any of those attempts produced a confirmed breach of a named production site. A production site is a live one serving real users.

Sam That's awkward for anyone arguing the public announcement caused everything, since traffic was already climbing. Still, the announcement clearly made it worse. What were operators who couldn't patch right away supposed to do in the meantime?

Maya The NSA and European agencies told them to take public PHP services offline. Here is a worked case. Say you run a government portal on an older PHP release. The patch means moving to a current version, which can take testing time you don't have. Leaving it up means anyone on the internet can try the exploit against it. Taking it down costs you an outage but closes the door.

Sam Plenty of people would shrug at that and assume nobody followed such advice. Did any real sites go dark?

Maya A string of them did, on 21 and 22 October: government portals, transport sites and banks. That is why the story reached consumer-facing news, not only security coverage.

Sam Why would so much old infrastructure sit exposed in the first place? A bug from more than a decade ago feels like something somebody should have caught by now.

Maya PHP has been the default language for a large share of the web since the mid-2000s. That includes the content-management systems, the software that lets non-programmers publish pages, still hosting blogs, shops, newsrooms and public-service portals. Memory-management choices made more than a decade ago turned out to be exploitable only when automated reasoning could chain the steps together. A human might see each step as harmless on its own.

Sam Faster clock for everyone, then. Defenders get speed, attackers get speed, and the people running the older servers are left deciding whether to keep the lights on.

Sam The second story today is a campaign ad that has people inside one party arguing with each other in public. Before we get to the reactions, I need the basics, because I've only seen the headline. What is the ad, and who is it aimed at?

Maya It's a 60-second spot from Ramaswamy, the Republican nominee for governor of Ohio, and it features the mother of his Democratic opponent, Acton. Her name is Arthur. In the ad she says Acton has lied about her past. Ramaswamy pledged 4 million dollars for it, enough, his campaign said, to keep it on air through election day. The two women haven't spoken in nearly fifty years.

Sam My guess is that the party closed ranks around its own nominee, the way parties usually do in the last weeks before an election. The ad might be rough, but nobody wants to undercut their candidate.

Maya It split instead. Senator Jon Husted, the Republican Senate nominee in Ohio, told the Associated Press, "I would not have run that ad." Nikki Haley, who called Ramaswamy "just scum" in 2024 after he brought her daughter into a primary debate, wrote on 20 October, "Just seeing what we tried to tell everyone back then. He's a bad guy." And Governor Mike DeWine, who has endorsed Ramaswamy, called it "deeply, deeply troubling" on 22 October.

Sam A governor who has endorsed the candidate is the surprising voice there. Did he withdraw that endorsement after saying it? And did he ask for the ad to come off the air, or only say he disliked it?

Maya He urged the campaign to take it down, and said his wife had seen it on television that morning. He also described the backstory: the indictment, the flight, the children never seeing their mother again. That made the ad "just more wrong," he said. But the endorsement stayed. In his words, "I'm not changing the endorsement. I endorse the whole ticket."

Sam Calling an ad wrong while still backing the man who ran it is a narrow line to walk. And the president took the opposite side the same day, which puts two of the party's most prominent figures against each other.

Maya Trump called it a "very powerful Ad" in a Truth Social post on 22 October and attacked DeWine as disloyal. He told reporters Ramaswamy had called him for advice and that he'd told him to leave it up. He said twice that he knew "nothing about" the childhood backstory, and called Acton "fair game." Vice President JD Vance, campaigning in Ohio, defended the ad as calling Acton a liar and told critics to stop whining.

Sam Everyone is arguing over one ad, which makes me wonder why it would carry so much weight. Is the race close enough that a single spot could change the result?

Maya The election is 3 November, twelve days after those remarks. The 270toWin polling average as of 22 October had Acton at 49.0 percent and Ramaswamy at 45.8. Individual polls range from a Marist survey at 50–43 for Acton to an ActiVote poll of 400 likely voters with Ramaswamy ahead 51–49. Several of those gaps sit inside or near the margin of error. For comparison, the Republican won Ohio's last governor's race in 2022 by 25 points.

Sam I'd like the one thing to take away from each of today's stories. Start with PHP2Shell, then the ad. Keep it to a single point for each, because I'll lose the thread otherwise.

Maya On PHP2Shell, the settled fact is that the surge in exploit attempts began before the official announcement and sped up afterward, while whether any named live site was breached was still open on 23 October. On the ad, the settled fact is that DeWine, Husted and Haley criticised it and Trump and Vance defended it, while whether it moves the polls was still open on 23 October.

Sam That's what I'll carry with me. This show is AI-generated, the voices are synthetic, and the facts come from the sources we'll name in a moment, whose authors have no affiliation with us. We're back with the next episode tomorrow, so thanks for listening, and see you then.

Maya Thank you for spending your time with us today, and we'll be here again tomorrow with the next two stories that everyone is talking about, so please come back and keep up with us.

Sam Everything today came from reporting by the Associated Press, The Washington Post, the Statehouse News Bureau, Politico, ABC News, NBC News, CNN, CBS, the Columbus Dispatch and 270toWin, along with polling from Marist, Fox News, Bowling Green State University with YouGov, and ActiVote, plus a Wikipedia compilation of polling averages. The PHP story drew on Wiz, CNN, Politico, TechCrunch, OpenAI, the PHP Foundation and Cloudflare.