The Defender's Window Nobody Planned For
For a few hours on Tuesday, the internet's most common programming language became its most dangerous. Two critical PHP flaws, tracked as CVE-2026-109618 and CVE-2026-109619 and nicknamed "PHP2Shell," let an unauthenticated attacker run commands directly on a web server. Both carry a CVSS v4.0 severity score of 9.2. The PHP Foundation shipped emergency fixes at 14:00 UTC in versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, and told every operator running the language to update at once. The urgency was not theoretical. Within nineteen minutes of the patch going live, the security firm watchTowr reported active exploitation attempts against both bugs. Cloudflare, which had switched on protective firewall rules by default hours earlier, said it was already seeing attack traffic aimed at its customers. PHP powers an estimated seventy percent of the world's websites, which means these critical PHP flaws reached everything from hobby blogs to the login pages of large enterprises. What makes this disclosure different from the parade of vulnerabilities that defines modern security is not the severity. It is the question of who got there first. For the first time in a headline breach of this scale, the flaws were surfaced by frontier artificial intelligence working for the defense, not discovered in the wild after the damage was already done.
A Bug Older Than Most Startups
The root of the problem is almost twenty years old. The weakness that became CVE-2026-109618 traces back to code written around 2004, and it surfaced again during an external audit of PHP in 2024. A core developer flagged the signs of a memory handling defect, but a working attack was only theorized at the time, and the fix stalled for months inside an organization that runs the backbone of the web on a shoestring budget. That delay became its own story. In an unusually candid statement, the PHP Foundation called the lag between discovery and remedy "a structural failure" of its decision process and its resources, and said it has since changed how it makes security decisions. It is a rare admission from an open-source steward, and a reminder that the plumbing of the internet often rests on a handful of volunteers and a thin balance sheet. These critical PHP flaws live in the PHP runtime itself, not in the applications built on top of it. That distinction matters enormously. A site can be written with flawless, security-conscious code and still be wide open, because the vulnerable logic executes before application code ever runs. On servers using the common mod_php configuration, a single malicious request can trigger the bug and let an attacker read the live memory of worker processes, potentially harvesting passwords, session tokens, API keys and TLS certificates belonging to entirely unrelated visitors.
The Machines That Went Looking
The discovery is where this episode turns into a landmark. Researchers at the cloud security firm Wiz, a startup called Calif, and OpenAI's Codex Security team each converged on the flaws, and they did it with help from the newest generation of AI models. Wiz says it used OpenAI's GPT-5.6 Sol alongside its own Wiz Atlas system and Google's Gemini 4 Argon to surface the first vulnerability on August 19. The work ran under OpenAI's "Patch the Planet" program, a defensive effort that pairs frontier models with human researchers to hunt bugs in widely used open-source software before criminals can. OpenAI framed the result as proof of a "Defender's Window," claiming that, to its knowledge, AI found the vulnerabilities before attackers did. Wiz Research then combined the two bugs into a single exploit it called php2shell and demonstrated unauthenticated remote code execution against the latest container images of WordPress, Drupal and Joomla, the three platforms that run much of the open web. One researcher reportedly compressed into roughly a week the kind of work that would normally occupy a team for months. The models are not magic, and the people involved were careful to say so. Calif noted that today's systems do not reliably find obscure, decades-old bugs entirely on their own, and that deep human knowledge of the target's history was still decisive. The lesson is less that AI replaced the researchers and more that it acted as a force multiplier, turning a faint signal from a 2024 audit into a fully weaponized proof of concept at a speed no human team could match. When Coordination Comes Apart If AI was the hero of the discovery, human coordination was the villain of the disclosure. The careful choreography meant to govern the release of critical PHP flaws like these unraveled in public. On Monday, a post on a popular forum leaked the two CVE identifiers and the intended disclosure date, forcing the PHP security team to pull its release forward by a day. Around the same time, a proof-of-concept video was posted by accident, accelerating the race to reproduce a working exploit. The defenders scrambled. Cloudflare enabled web application firewall rules by default at 03:02 UTC, hours before the official patch, and the container specialist Chainguard published fixed images roughly six hours ahead of the source release. "We are seeing early attacks emerging and are updating mitigations as new vectors show up," said Cloudflare chief technology officer Dane Knecht, who urged administrators to patch rather than lean on filters. Attackers, predictably, began mutating their payloads to slip past the new rules almost as fast as they were deployed. The fallout spread beyond server logs. The HR software provider BambooHR took services offline across the United States, Canada and Ireland as a precaution, and several Dutch government and education sites did the same. The Netherlands' national cyber-security center said it had not observed the flaw being abused against the systems it monitors, a small mercy in an otherwise frantic day. What Every Operator Should Do Now The practical advice is blunt. "The only thing I can say is that the only mitigation is patching or upgrading PHP," said core developer Gina Peter Banyard. Firewall rules and configuration tweaks reduce exposure, but only the updated runtime closes the door completely. Anyone on a supported branch should move to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 without delay, and check distribution changelogs, since many vendors backport fixes under their own version numbers. The harder conversation is about the long tail. PHP 8.1 and everything older is already unsupported, and PHP 8.2 reaches end of life on December 31, 2026. Source backports exist for 7.4, 8.0 and 8.1, but a vast population of neglected servers will never be touched, leaving these critical PHP flaws as a permanent fixture of the threat landscape for years. The PHP Foundation can ship a patch. It cannot force the world to apply it. Step back, and the episode reads like a preview of the next decade of security. Frontier models can now compress the discovery of a dangerous flaw from months into days, which is reassuring when they wear a white hat and alarming when they do not. The same capability that handed defenders their window will, sooner or later, hand attackers one too, and the leaked disclosure shows just how narrow that advantage can be. For a wider view of how autonomous systems are reshaping the attack surface, see Blumefield's reporting on AI-driven cyber operations. The uncomfortable truth behind these critical PHP flaws is that the contest between machine-speed offense and machine-speed defense has already begun, and the human institutions caught in the middle are the slowest part of the system. Follow Blumefield for continuing coverage.