PHP2Shell Shows How AI Is Finding Vulnerabilities Quicker Than We Can Handle
AI DIGEST · NOVEMBER 3, 2026
AI-powered cyberattacks raise fears about the future of cybersecurity
On Oct. 20, 2026, the identifiers for two critical security flaws in PHP began circulating online. PHP is an open source scripting language used by roughly 70% of websites whose server-side language W3Techs can identify. The flaws, collectively known as PHP2Shell, each received a severity score of 9.2 out of 10. Both can let an attacker run code on an affected server without signing in, putting sensitive data and infrastructure at risk. AI tools helped researchers identify and test the issues in days. (, , )
The vulnerabilities are in PHP itself, so a website could be exposed even if its own application code had no flaws. Wiz says affected releases go back to PHP 5.4, which came out more than a decade ago. PHP released security updates on Oct. 21, and the U.S. Cybersecurity and Infrastructure Security Agency urged organizations to patch immediately. (, )
From flaw identification to a working exploit
Wiz attributes the speed of its research to Google’s Gemini 4 Argon, OpenAI’s GPT-5.6 Sol, and its own Wiz Atlas system. The researchers used these tools to find bugs, develop and test exploits, and validate the patch. In controlled tests, they demonstrated attacks against WordPress, Drupal, and Joomla, three platforms that run on PHP. ()
Wiz found its first vulnerability on Aug. 19 and reported its findings to the PHP security team on Aug. 24. PHP developers had noticed signs of the underlying memory-handling problem in 2024, but they had yet to establish a concrete way to exploit it. The PHP Foundation says fixing it became a top priority on Aug. 11, when Calif researcher Ryan Austin reported that he had used AI-assisted research to demonstrate remote code execution. The patches arrived more than two months later. (, )
The disclosure timeline
PHP had planned a coordinated release for Oct. 22. The initial leak on Oct. 20 contained no technical details, but it pushed maintainers to move the release forward by a day. Later that evening, Calif mistakenly published a demonstration video. Early on Oct. 21, Chainguard published container images containing the fix more than six hours ahead of the coordinated release. That made the patched software available to examine before most users could update. ()
Amazon said it detected an attempted attack on Oct. 17 against its honeypot network, a set of decoy systems built to attract attackers. It could not tell how the attacker had learned of the flaw or whether the attempt would have worked on a real system. ()
Sophos initially reported a successful break-in, then corrected its assessment on Oct. 22 to say it had observed scanning and a small number of failed attempts. On Oct. 23, the Netherlands’ National Cyber Security Centre said exploitation appeared complex and that it had one unconfirmed report of possible exploitation abroad. The centre and several other organizations had temporarily taken their sites offline as a precaution. (, )
Later reporting established that the flaws were being exploited in specific environments. Palo Alto Networks reported active exploitation targeting internet-exposed PAN-OS management interfaces. Dream Research documented one campaign that compromised 25 firewall and management servers by combining PHP2Shell with a separate vulnerability. The full scale of compromise remains unclear. (, )
The gap between finding a flaw and fixing it is widening
A central concern is that attackers can use the same advances in AI that helped researchers develop these exploits. Calif says its initial discovery used generally available models, and the work still depended on the researcher’s knowledge of PHP and its vulnerability history. As it takes less time and effort to investigate complex flaws, defenders have to assume that others can examine the same code with similar tools. (, )
For an organization that has already been compromised, installing a patch may leave other problems unresolved. An attacker could retain access through a backdoor or stolen credentials. The UK’s National Cyber Security Centre advised organizations unable to update or take affected systems offline immediately to assess for signs of compromise, including backdoors. That guidance addresses the risk of continued access. How often that happened in this case remains unknown. ()
In this case, AI helped compress complex vulnerability research into days, according to Wiz, while producing and coordinating a fix took months. Releasing that fix still left organizations responsible for finding and updating affected systems. The gap between discovery and protection gives attackers an opportunity that faster research alone cannot close. (, )
What AI means for the future of cybersecurity
Financial institutions, hospitals, universities, and governments depend on software whose foundations were written decades ago. Small teams with limited resources maintain some of those dependencies. The PHP Foundation acknowledged that scarce expertise and difficulties making decisions had caused earlier work on this problem to stall. Familiarity with old code offers little assurance about flaws that have never been found. As AI lowers the cost of searching for them, institutions that depend on that code will need to fund its maintenance with the same seriousness they apply to new products. ()