AI Found PHP's Most Severe Vulnerabilities in a Decade in Two Days
Production PHP servers are exposed to unauthenticated remote code execution and cross-request credential theft, and the affected version range is so wide that roughly 60% of public-facing PHP servers are already unsupported. The two-day AI-assisted discovery also compresses the window defenders have to react: a vulnerability class that used to surface over months can now be weaponized before a patch is ready.
Two critical memory safety bugs in PHP core now have working exploit chains, and the discovery was done in two days by a Wiz Research researcher steering two AI models. CVE-2026-109618 and CVE-2026-109619 both score 9.2 on CVSS v4.0, require no authentication, and allow remote code execution. The root cause sits in PHP's custom zalloc heap allocator and its request lifecycle: OOM interruptions and fatal-error paths leave internal structures uninitialized, and subsequent requests can read leftover memory, including other sessions' tokens, API keys, and TLS private keys.
Traditional discovery for this class of bug has often taken months, and 40% of 47 critical PHP vulnerabilities disclosed between 2015 and 2020 went unnoticed for over a year. The AI-assisted workflow split the work: Gemini 4 Argon traced how external data reaches memory allocation code, GPT-5.6 Sol modeled behavior and generated PoC candidates, and a human researcher verified. Three iterations produced reproducible attack chains for both CVEs.
Because PHP still powers about 70% of active websites and PHP2Shell claims compatibility with every version from 5.4.0 onward, roughly 60% of public-facing PHP servers are already unsupported. Cloudflare shipped emergency WAF rules, Elastic and Corelight published Suricata/Zeek detections, and the PHP Foundation moved up a patch that took about two months to prepare.
AI compressed the information-gathering phase of vulnerability research from weeks to hours, but the entry point still came from a human prompt and the output needed human validation; the speedup is real, but it is not autonomy.
The same two-day discovery timeline is a warning for operators, not only a research milestone: once a serious flaw can be found and weaponized that fast, the long tail of unsupported PHP versions becomes a much larger liability.
The vulnerabilities are a product of PHP's backward-compatibility debt; memory reuse and preserved 'living fossil' session behavior made the bug class hard to find with traditional fuzzing, and the patch addresses the symptom while governance and compatibility debt remain.
A wrong prompt can send the model off on irrelevant problems, which keeps the human researcher in the loop as the one who frames the problem; AI narrows the search space, it does not define it.
The funding model for PHP core security is now the bottleneck: the PHP Foundation has a limited team, money flows from the WordPress/Drupal/Magento ecosystem are opaque, and a two-month patch cycle is too slow once AI cuts discovery time to days.