FULL STORY

OpenAI Model Uncovers Critical PHP Vulnerabilities

OpenAI's GPT-5.6 uncovered critical pre-auth RCE flaws in PHP, triggering an emergency patch affecting most of the web. POCs proved viable and researchers expect LLM-driven code audits to surface dozens more.

2026-10-21 ~ 2026-10-22 · 5 episodes · 58 posts

Episode 1 · PHP's php2shell: Two Unauthenticated RCE Flaws Found with Help from OpenAI Models (2026-10-21, 49 posts)

PHP shipped an emergency security update on October 21 at 14:00 UTC, patching two critical vulnerabilities allowing unauthenticated remote code execution (RCE)—CVE-2026-109618 and CVE-2026-109619 (codenamed php2shell)—in versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35. All PHP deployments handling web requests are affected. OpenAI announced its models played a key role in the discovery and launched the "Patch the Planet" global cybersecurity initiative; the FBI, NSA, and other agencies also issued emergency alerts. Since PHP powers roughly 70% of websites worldwide, this is regarded as one of the largest-impact open source security incidents in recent years.Confirmed

  • PHP's official security advisory confirms the two vulnerabilities allow unauthenticated RCE, affecting all web-serving deployments including PHP-FPM, modphp, php-cgi, FrankenPHP, and LiteSpeed; modphp carries the highest risk, with a demonstrated exploit triggerable by an unauthenticated request before application code runs (m2, m13).
  • Both are rated CVSSv4 9.2; security researcher cyb3rops revealed one is a core RCE flaw scoring a perfect CVSS 10.0 (m11, m20).
  • The vulnerabilities were found and disclosed collaboratively by Calif.io, OpenAI, Wiz, and Quarkslab; OpenAI says researchers in its "Patch the Planet" program made the discovery using OpenAI models, and over the past two months led a coalition of maintainers and security vendors to prepare the fix (m1, m16, m18, m19).
  • The flaws lie in PHP itself, not plugins—plugin developers cannot fix them, and sites like WordPress must upgrade PHP directly (m6).
  • The FBI published both CVE numbers and advised organizations unable to upgrade immediately to take systems offline; the NSA also issued an immediately effective emergency recommendation to update or take systems offline until fully mitigated (m12, m14).
  • Cloudflare has urgently shipped WAF rules as a temporary mitigation (m15).
  • The joint write-up mentioned in m16 uses a "journey through time" narrative tracing the root cause back through code evolution since 2004.Unconfirmed
  • vxunderground calling the flaw "ultra mega" severity on par with Log4Shell is an analogy from some security researchers, not an official rating (m17).
  • m11 states that older versions (pre-PHP 8) are expected to receive no patches, consistent with the advisory listing only four PHP 8.x fixed versions, but officials have not explicitly stated the plan for older versions.
  • m4's mention of "AI-assisted attacks possibly arriving within 15 minutes" is a precautionary speculation from the Japanese security community; no real-world exploitation has been confirmed.Why it matters
  • PHP powers about 70% of websites worldwide; the flaws are low-barrier and automatable, and because they sit in the language core rather than the plugin layer, the fix falls on server operators—WordPress site owners were urged to stand by overnight to patch.
  • This is a landmark case of AI models systematically participating in vulnerability hunting for critical open source infrastructure; OpenAI leveraged it to launch the "Patch the Planet" program, signaling the normalization of AI-driven security research.
  • The rare simultaneous warnings from agencies in multiple countries (FBI, NSA) and Cloudflare's emergency WAF rules reflect the incident's potentially huge impact.

29 more related posts →

Episode 2 · Researchers Predict LLM Audits Will Uncover More PHP Vulnerabilities (2026-10-21, 3 posts)

Security researcher moyix predicts researchers will point LLMs at PHP codebases and uncover dozens of additional serious vulnerabilities beyond upcoming fixes, betting on an exploitable stack overflow in recursive abbreviation-handling code.

Episode 3 · PHP RCE proof-of-concept confirmed working, but not one-click (2026-10-21, 2 posts)

Security researchers confirmed the leaked PHP RCE proof-of-concept is genuine and usable, though exploitation requires certain preconditions. Combined with the vulnerability's high severity rating, the situation remains concerning.

Episode 4 · POC Released for Critical Pre-Auth RCE in PHP session.upload_progress (2026-10-21, 2 posts)

A researcher disclosed two critical pre-auth RCE flaws (CVSS 9.2) in PHP's session.uploadprogress handler via the php2shell project, and the POC has leaked early. Researchers warn mitigations alone may not suffice and urge exposed servers to patch promptly.

Episode 5 · OpenAI Says PHP Vulnerability Finds Are Just the Beginning as Devs Joke Online (2026-10-21, 2 posts)

After OpenAI announced that GPT-5.6 (Sol) uncovered critical PHP vulnerabilities, the company said more discoveries are yet to come, while developers joked online about the 'PHP archaeologists' finally facing their reckoning.