Amazon says two PHP vulnerabilities, found with OpenAI and Google AI models, were exploited four days before public disclosure and are now broadly exploited; PHP releases patches

PHP maintainers have released patches for two serious security flaws, nicknamed PHP2Shell, that went unnoticed for over a decade. Researchers at Calif and Wiz found them with help from OpenAI and Google AI models, and OpenAI helped verify them and develop the patches. Amazon says it saw an exploitation attempt four days before public disclosure and that the flaws are now broadly exploited. The NSA and UK NCSC have told organisations to update immediately or take systems offline. This comes second-hand: the Ars Technica piece draws on FT reporting, and no CVE IDs, affected versions or advisories are linked.

Anyone running PHP on a public server, including WordPress sites and custom applications, should check their version and patch now. The flaws reportedly allow server takeover without credentials, and active exploitation means delay carries real risk. Security teams should also check logs for signs of compromise from before the patch, since attacks started ahead of disclosure. For AI, it is a concrete example of models finding old, widely deployed bugs, but the same capability can speed up attackers' exploit building. How much of the discovery was the AI and how much was human expertise is not clear from the coverage.

The "about 70% of websites" figure is Wiz's own estimate of PHP use. It does not mean 70% of sites are vulnerable, because that depends on version and configuration. "One of the most impactful security issues in recent memory" is OpenAI's own description, and OpenAI helped find and fix the flaws, so treat it as an interested party's claim. The reports of European governments taking sites offline are described as unconfirmed, and the AI-capability framing says little about how much of the discovery was automated.

1. 01Official PHP advisories and CVE entries confirming the affected versions, severity scores and fixed releases, which would replace the current second-hand detail.
2. 02Independent exploitation data, such as CISA's Known Exploited Vulnerabilities listing or telemetry from security vendors showing scan and attack volumes, to test the claim of broad exploitation.
3. 03Technical write-ups from Calif, Wiz or OpenAI that show how much the AI models contributed to the discovery, and whether other AI-assisted finds of similar scale follow.