AI Helps Expose Critical PHP Flaws

AI-assisted security research has helped expose two critical vulnerabilities in PHP, the software behind many websites and web applications. Security updates arrived on October 21. The urgent task now is getting those fixes onto the systems that need them.

Today’s episode leads with the PHP story, then examines a developer’s cloud-bill allegation, Railway’s credential warning and an AI modding tool for Fallout 2.

Find AI World Today: Daily News on Spotify, Apple Podcasts and Amazon Music. Browse more episodes in the AI World Today podcast archive.

The vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619, affect PHP’s handling of memory when requests fail. Under certain conditions, an attacker can use them to execute code on a server without logging in.

PHP’s official advisory warns that installations serving web requests are at risk. Apache with mod_php is a particularly serious case: an attack against one flaw can happen before the website’s own application code runs.

Cloudflare’s technical explanation makes the practical distinction clear. Updating a site’s plugins or application code alone does not repair the underlying PHP runtime. Website owners whose hosts manage PHP should ask those providers to confirm that the security update is running.

The supported branches received these fixed releases:

Some distribution and support vendors provide backported fixes. Check the vendor’s advisory when assessing an older-looking package version.

OpenAI says Calif identified a critical flaw using its models, while Wiz independently discovered and reported both vulnerabilities. Google says Gemini contributed to finding the problem, developing test attacks and validating patches.

The PHP Foundation’s account adds important history. Warning signs around one issue appeared in 2024, before researchers had established a concrete route to remote code execution. Compatibility concerns, limited specialist resources and slow decisions complicated the work. Reports in August 2026 demonstrated the severity.

The Foundation acknowledges a structural failure in its process. Its account also makes clear that humans continue to make security decisions and work on the patches.

The implication is broader than this release: AI can accelerate investigation, while maintainers and infrastructure operators still carry the work of turning findings into protection.

Reuters reports that Cloudflare’s chief executive described attempted exploitation rising into the millions after disclosure. That is an attempt count, not a count of successfully breached websites.

Amazon says its decoy-server network detected an attempted attack on October 17, before public disclosure. It cautions that the early attempts’ effectiveness against real systems is unclear.

AI World Today reports that developer Just Lingonberry showed $88,982.85 in estimated Google Cloud charges and blamed a Codex-assisted deployment for exposing credentials. The explanation relies partly on another model’s analysis and remains unverified.

The developer also shared a message saying Google had cleared the balance. Today’s episode includes that reported adjustment.

For builders, Google’s documentation reinforces a useful distinction: spending caps cover eligible services, while standard budget alerts alone do not stop usage. Even cap enforcement can be delayed.

In its October 21 disclosure, Railway said an unauthorized party may have accessed encrypted customer environment variables and obtained the decrypted contents of some of them. It recommended that all users replace stored secrets and review activity.

Its October 22 guides explain how to check Google Cloud and Amazon Web Services for misuse. Revoking exposed credentials at the issuing service is part of the response; changing a value inside an application alone does not invalidate the old key.

The lighter closing story is Rerun Reactor, an unofficial fan project connecting AI assistants to Fallout 2 modding tools.

Its guide describes changes to items, dialogue, rewards and companions. A player describes an idea, and the assistant assembles a mod recipe that can be added to the browser game.

Players must supply their own game files. The assistant can check the recipe, but it cannot playtest the result. Human judgment still determines whether the change works and makes the game more enjoyable.

For returning players, that could make a small personal quest or companion a compelling reason to revisit a familiar world.

Follow AI World Today: Daily News for the next edition. This episode is narrated by an AI voice.