PHP's October 21 security release warrants operational priority: two critical vulnerabilities can allow unauthenticated remote code execution. The project urges immediate updates to 8.5.12, 8.4.27, 8.3.36 or 8.2.35, depending on the branch in use.
The discovery also offers a concrete case of AI-assisted security research. OpenAI credits Calif with identifying one flaw using GPT-5.6 Sol. Wiz independently discovered and reported both, using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol in its research. PHP's maintainers led the fixes and releases.
The failure can precede application code
CVE-2026-109618 concerns out-of-memory handling. An interruption can leave internal structures partly initialized, allowing request shutdown to process invalid data. Researchers demonstrated exploitation against Apache mod_php before application code begins executing.
CVE-2026-109619 involves references to freed memory surviving between requests handled by the same worker. A later operation can consume that stale state before normal initialization replaces it. Exploit conditions vary with configuration and execution path.
Demonstrating an attack and assessing the full exposure are different tasks. The first advisory considers PHP-FPM susceptible but records no demonstrated complete exploit for that interface. The maintainer's update guidance includes FPM, mod_php, php-cgi, FrankenPHP and LiteSpeed. Operators need to assess their installations without treating an absent demonstration as proof of immunity.
That places remediation below the application layer. Updating a framework, reviewing application permissions or requiring login cannot, on their own, fix an interpreter flaw that may occur before those protections run.
What AI changes in security work
Wiz says frontier models helped develop and adapt exploits across PHP versions, configurations and memory layouts. That is the participating researchers' account, without an independent measurement of the human effort saved. It nevertheless describes a more specific use than asking AI to find a bug: investigating failures, testing hypotheses and refining a complex technical chain.
The same capability can reduce attackers' effort. Cloudflare has already observed attack traffic and changing payloads intended to evade its protections. It has deployed WAF rules but continues to recommend patching.
MnzAI Labs' assessment is that this puts more pressure on engineering response time. A stronger discovery tool loses some of its defensive value if an organization takes too long to establish where it uses the component or who can update it. The AI discussion therefore needs to reach software inventories, service ownership and the ability to distribute fixes.
The fix must reach the process serving traffic
Verification can start in the wrong place: command-line PHP is not necessarily the installation used by FPM or Apache. The server-checking guidance calls for confirming the serving runtime and full package revision. Distributions may backport security fixes without adopting the upstream release number.
That also means examining container images and services managed by other organizations. Installing a package is one step; verifying that running processes received the update and the application still works completes the task.
The php2shell-checker can help identify versions, but it compares information from HTTP headers or the local binary. It does not independently establish that the fix is present, and an unknown result calls for further investigation. PHP's temporary mitigations also offer only partial protection.
For a production team, the useful evidence at change closure is concrete: the vendor advisory, the installed package, the updated processes and application test results. That is the path from AI-assisted discovery to protection for users.