AI safety
AI helped find two old flaws in PHP, which runs much of the web. The US now lists them as exploited.
Published October 27, 2026
Security researchers working with AI models found two critical flaws in PHP, the programming language behind WordPress, Drupal and many other websites. Some of the weak code had been there for more than a decade. PHP released fixes on October 21, and on October 26 the US cybersecurity agency CISA added both flaws to its list of vulnerabilities known to be exploited in real attacks.
What happened
The flaws, nicknamed PHP2Shell, allow remote code execution: an attacker on the internet can make a web server run their own commands, with no login. They are memory-safety bugs. When PHP hits its memory limit at just the wrong moment, it can leave half-cleaned-up data behind, and the next visitor’s request may reuse it. PHP rates them 9.2 out of 10 on CVSS v4.0, the current version of the standard severity scale. Wiz, one of the finders, says it also showed attackers could read passwords and keys out of a server’s memory.
Three teams found the problem. The security firm Calif began auditing PHP on July 30 under OpenAI’s Patch the Planet program and reported the first bug on August 11; OpenAI independently reported the same root cause, and says Calif used its GPT‑5.6 Sol model. Wiz found it separately on August 19 using Gemini 4 Argon, GPT‑5.6 Sol and its own Wiz Atlas tool. Calif’s researcher, Ryan Austin, writes that the AI helper was steered toward a 2004 PHP bug, found by Stefan Esser, that the author already knew well.
The PHP Foundation says the problem was not entirely new. A 2024 outside audit turned up early signs of it, but nobody could prove it was exploitable, and work on a fix stalled for months. The Foundation calls that delay “a structural failure.” The AI-assisted reports made the danger plain.
The release was also rushed. The plan, called coordinated disclosure, was to keep details secret until the fix shipped. On October 20 the bug IDs leaked and Calif mistakenly posted a demo video; early on October 21, the software distributor Chainguard published fixed builds more than six hours ahead of the release, the Foundation says. PHP moved the release up a day.
Why it matters to a normal person
If you run a website, blog or online shop on PHP, the fix is to update PHP to 8.5.12, 8.4.27, 8.3.36 or 8.2.35, or confirm your hosting company has. PHP says there is no complete workaround, and PHP 8.1 and older no longer get official updates. CISA tells organizations that cannot patch to take those systems offline.
Our advice: if you only visit websites, there is nothing to install; the risk is to data those sites hold about you.
The bigger story is speed. Wiz says a single researcher with frontier models did in a week what once took a world-class team months. OpenAI says it expects findings like this to “increase sharply in the coming months.” The PHP Foundation says AI-assisted research has brought far more reports, including more than 100 since July 1 that were invalid, duplicates or not reproducible, to a small group of PHP developers.
What it does not mean
It does not mean an AI found this on its own. In the finders’ write-ups, people chose the target and guided the models. Calif’s researcher writes that no model can unearth “all bugs” with simple prompts.
It does not mean every PHP site has been hacked. InfoQ reports that Amazon saw one exploitation attempt dating to October 17, and BleepingComputer says Cloudflare saw early attacks before the patch. But the security firm Sansec, quoted by InfoQ, says the public exploits do not work against a default PHP installation. InfoQ also reports, citing JFrog, that fake exploit code is circulating.
And the big-impact language is the finders’ own. The “70% of the internet” figure and “arguably in software history” come from Wiz, which sells security products. One smaller related issue is still unpatched; PHP calls it unlikely in practice and suggests turning off a feature called session upload progress until its next release.
Related: A Cloudflare AI model sent some users pieces of other customers’ data. Here is how it happened., Anthropic says small AI models may soon copy themselves on home gadgets
Sources
- Security advisory: Critical remote code execution vulnerabilities, update now — PHP project (primary source; Oct 21 and Oct 23, 2026)
- Today’s PHP security release, and how we’re strengthening PHP security — The PHP Foundation (primary source; Oct 21, 2026)
- PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet — Wiz Research (finder; Oct 21, 2026)
- PHP2Shell: A tale of two PHP remote vulnerabilities — Calif (finder; Oct 21, 2026)
- Patch the Planet: global response coordination as agents uncover critical vulnerabilities — OpenAI (Oct 21, 2026)
- Critical Vulnerabilities in PHP Require Immediate Patching — CISA (government alert; updated Oct 26, 2026)
- PHP ships emergency patches for critical PHP2Shell flaws — BleepingComputer (independent report; Oct 21, 2026)
- PHP Patches Two Critical RCE Flaws That Were Already under Attack — InfoQ (independent report; Oct 25, 2026)