2026-10-24 · contested story

The PHP Meltdown Shows How AI Makes Bugs More Exploitable

On October 21, 2026, the PHP Foundation released emergency patches for two critical vulnerabilities (CVE-2026-109618 and CVE-2026-109619, both rated 9.2 CVSS), dubbed 'PHP2Shell,' affecting a programming language that powers roughly 70% of the world's websites. The flaws—unauthenticated remote code execution bugs that had lurked in PHP's core for over a decade—were discovered using frontier AI models: security firm Calif found one using OpenAI's GPT-5.6 Sol as part of OpenAI's 'Patch the Planet' initiative, while Google-owned Wiz independently found them using Gemini 4 Argon, GPT-5.6 Sol, and its Wiz Atlas tool. Governments from Finland to Japan to Canada took websites offline as a precaution, and the central narrative tension is that AI both uncovered these flaws before attackers and dramatically lowers the barrier for attackers to exploit them—what nearly every outlet calls a 'double-edged sword' or 'AI inflection point.'

How each side frames it

left
"The concern is that every hacker will soon have access to a model as strong as Claude Mythos, and will be able to target any institution that is slightly behind on its cybersecurity—from your local utility company to the hospital down the road."
"AI is already putting us in a world with more possibilities, and more dangers. There's a whole lot of work to do to keep up."
"What the US president exhibits is a fundamental incompetence that places us all in grave danger."
"Current and former Anthropic researchers have said publicly that these systems could kill off humanity within the decade."
"'There have to be guardrails and safety measures in place which are way ahead of the capability which is being developed,' said O'Shea."
center
"It highlights how AI can uncover security flaws with internet-scale impact, but the same capabilities can also help attackers rapidly develop offensive tools and exploits."
"We believe PHP2Shell could mark a similar turning point for the AI era, illustrating how AI can change the practical exploitability of longstanding vulnerabilities."
"This changes the economics of both discovery and exploitation. As AI reduces the effort needed to develop and adapt complex exploits, vulnerabilities that once seemed impractical to exploit become viable targets."
""Defenders have always relied on the gap between 'vulnerable' and 'exploitable' to buy time to patch," Michael Leland, field CTO at Island, said in an email to Axios. "That gap is rapidly closing.""
""This is being touted as like, 'Hey, you can push a button and basically take over a server type of thing,'" he said. "…To actually pull this off requires a couple of conditions that are actually fairly complex and not always going to work out in favor of somebody who's trying to go and exploit this bug.""
right
"It is equally a case study of the importance of open security research, and a cautionary tale about what can happen if vulnerability disclosure is concealed or even weaponized."
"paraphrased: Cybersecurity stocks retreated Friday amid investor fears over AI 'disruption' from Anthropic's Claude tools."
"paraphrased: Democratic governors are one-upping each other with costly regulations that could cripple America's AI buildout."
"Every second of every day we're at war with these cyberterrorists, and it's time public policy reflected this."
"Too many municipalities remain unprepared for today's threat-environment, with inconsistent software updates, weak IT departments, and a pattern of selecting the insurer-paid option when confronted with the cost of restoring systems."

What each side left out

The right left out — covered by the CNN & Wiz & OpenAI & The Guardian & The Atlantic
The right left out — covered by the Common Dreams
The left left out — covered by the Rapid7 & Cisco Talos & Qualys & Aikido
The left left out — covered by the Investor's Business Daily
The center left out — covered by the Common Dreams
The center left out — covered by the The Atlantic

What's actually true?

[verified] PHP powers roughly 70% of the world's websites.
[verified] The two vulnerabilities are CVE-2026-109618 and CVE-2026-109619, each rated 9.2 on CVSS v4.0.
[verified] The underlying flaws existed in PHP for over a decade (back to PHP 5.4/2004) but were considered too difficult to exploit until AI lowered the barrier.
[verified] The flaws were discovered by Calif using OpenAI's GPT-5.6 Sol and independently by Wiz using Gemini 4 Argon, GPT-5.6 Sol, and Wiz Atlas.
[verified] Government websites in Finland, France, the Netherlands, Germany, Canada, and elsewhere were temporarily taken offline as a precaution.
[verified] The patch details leaked ahead of the scheduled release, prompting the PHP team to move up disclosure and sites to go offline early.
[verified] Amazon detected an exploitation attempt against a honeypot/decoy system on Oct. 17, four days before public disclosure.
[contested] Reliable, broadly portable exploitation of the flaws was still difficult as of disclosure; most observed activity was scanning and reconnaissance.

The narrative clash

Whether the flaws were discovered with frontier/exclusive AI or publicly available models
Left: The researcher who found the weakness, however, said they used models available to the public.
Right: paraphrased: OpenAI said it led a coalition and that 'frontier AI' had been key to discovery, touting the role its model played.
Whether active in-the-wild exploitation was occurring at disclosure
Left: The primary U.S. cybersecurity agency, the Cybersecurity and Infrastructure Security Agency, said... it had 'not observed exploitation of these vulnerabilities as of the time of publication.'
Right: "PHP2Shell is now being broadly exploited in the wild," said Amazon.
How easy exploitation actually is
Left: the vulnerabilities can be exploited 'fairly easily' (Wiz, via WIRED); 'AI makes exploiting such vulnerabilities of immediate importance. Like, indeed, a minute or so.'
Right: Our attempts so far have produced a working exploit only under narrow lab conditions, and exploitation depends on assumptions that are unlikely to hold on internet-facing instances.

103 sources analyzed

Mother Jones CNN wiz.io Reason Fox News Mother Jones NPR Axios Boston Herald City Journal Common Dreams WIRED Techzine Global Reason National Review Common Dreams Los Angeles Times Palo Alto Networks Investor's Business Daily The Federalist Mother Jones The Washington Post CNA New York Post City Journal Common Dreams NPR CNET New York Post The Daily Signal Salon The Verge CBS News Reason The Federalist Mother Jones NPR TechCrunch New York Post Fox News Salon The Guardian ABC News - Breaking News, Latest News and Videos New York Post Fox News Mother Jones CNN CyberScoop Investor's Business Daily City Journal Common Dreams The Washington Post The Conversation Boston Herald Fox News Salon NBC News Wikimedia.org Investor's Business Daily The Federalist Salon PBS NewsHour OpenAI Boston Herald City Journal PBS NewsHour Cisco Talos Blog Investor's Business Daily The Federalist The Washington Post Rapid7 Reason National Review The New York Times Coinfomania National Review Los Angeles Times PCWorld National Review Global News Dunya News Engadget Google Cloud The Atlantic Bing Search NPR Rescana The Washington Post Mezha.net CNN OX Security U.S. News & World Report Qualys CNN PC Gamer Los Angeles Times SecurityWeek The Washington Post Yahoo CBC Aikido Security The New York Times Financial Times

Perspectivity reads the coverage — not just who published it.

Read the full interactive analysis →