Tomer Schwartz, Wiz's CTO of Exposure Management and AppSec, says the PHP flaw let attackers take over unpatched servers remotely without a password and had persisted for twelve years in default configurations.
Schwartz says leaked details forced the PHP Foundation and partners to accelerate their timeline, and many sites went offline for 48 hours before patches arrived.
Schwartz argues AI lets researchers do 100x or 1000x the work, and predicts more critical vulnerabilities will be published in coming months, leaving a long tail of unpatched systems exposed.
AI-found PHP flaw signals a coming flood of critical bugs
AI has made finding exploitable bugs in old, widely used code fast and cheap, and PHP two shell is the first big example. A twelve-year-old flaw in software running most of the web became a working remote exploit and leaked before patches shipped. Schwartz expects many more such disclosures, while most systems still patch too slowly to keep up.
Takeaways
Twelve-year-old PHP flaw exposed much of the web. PHP two shell let attackers take over unpatched servers remotely without a password, even in default configurations. PHP underpins WordPress, Drupal, Joomla and MediaWiki, roughly 70 percent of the web-facing internet. Banks, hospitals, retailers and government sites pulled themselves offline to patch.
AI is turning old bug hints into exploits. Similar problems were reported as early as 2020, but nobody had a working remote exploit. AI lets defenders test far more ideas against sensitive code, so Schwartz says the one-headline-bug-a-year cadence is ending. He predicts more critical disclosures within weeks or months.
Project Atlas multiplies research output 100x to 1000x. Wiz's custom harness combines frontier models from several labs and swaps them as models improve. Without AI, Schwartz says, finding these bugs might have taken months or a year. Wiz tested open weight models and found them extremely effective, though Atlas uses none today.
Most workloads remain unpatched, with a long tail. Cloudflare saw millions of exploit attempts. Solo developers and small IT teams lag, as do routers and firewalls that bundle PHP. Major hosting and website-building platforms patch behind the scenes, but small businesses running old WordPress sites are likely exposed.
Startups should automate patching from day one. Schwartz says manual patching cycles built before AI will not scale with the coming volume of vulnerabilities. Infrastructure should be code-defined and patched automatically. Trust-based businesses should bring security in house, perhaps among the first hires, or every week becomes a fire drill.
How PHP two shell reached the public
Calif found the first CVE in a security audit and reported it to OpenAI. Wiz found related bugs, and all parties reported to the PHP team, which opened a shared Slack channel staffed at every hour, weekends included.
The PHP team spent weeks building and testing patches that had to stay backwards compatible with PHP extensions. Detection to fix took about two months.
Early notice went to choke points like cloud providers and web application firewall vendors so they could ship detection rules. Linux distributors also needed time, since a php.net release does not mean distributions carry the fix.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
The circle of insiders grew toward hundreds and details leaked, so the timeline was accelerated and many sites went dark for 48 hours. Amazon saw exploitation about a week before patches. Schwartz blames no one, calling leaks a function of coordinated disclosure.