AI and software security: PHP's emergency fix and Claude's machine-checked proofs

Two pieces of security news this week show both sides of what AI is doing to software. On 21 October, the PHP project released emergency fixes for two critical flaws that let attackers run their own code on a web server. The PHP Foundation says AI-assisted research has sharply increased the number of bug reports it handles. On 22 October, Anthropic published research showing Claude writing mathematical proofs that hundreds of pieces of open-source code cannot crash. Finding bugs is getting cheaper. Proving there are none may be starting to.

What happened with PHP

PHP is the programming language behind WordPress and a large share of the websites on shared hosting. On 21 October the project released PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. They fix two vulnerabilities:

• CVE-2026-109618, remote code execution through memory corruption when PHP runs out of memory.
• CVE-2026-109619, remote code execution through stale global state, a "use-after-free" bug.

Both are rated 9.2 out of 10 on the CVSS severity scale. A score of 10.0 circulated before the release, but the PHP Foundation says that figure was wrong. The flaws can be triggered by a request from someone who is not logged in, before the website's own code runs. Attacks were demonstrated against Apache with mod_php using default settings, which the project lists as the highest-risk setup. PHP-FPM, php-cgi, FrankenPHP and LiteSpeed are also listed as affected.

The advice from php.net is blunt: there are no complete mitigations, and updating is the only safe choice. PHP 8.1 and older no longer get releases from the project, although minimal source patches were published for 7.4, 8.0 and 8.1 for anyone who builds PHP themselves.

How the bugs were found

The PHP Foundation's account is a good picture of security work in 2026. The out-of-memory issue was first glimpsed in 2024, while the team worked through findings from an earlier audit by Quarkslab, but the work stalled. On 11 August 2026, Ryan A of Calif reported a way to get reliable code execution. OpenAI independently reported the same root cause, and the release credits OpenAI's Codex Security. Further ways to trigger the bugs came from Wiz on 24 August and from OpenAI on 18 September.

AI also appears throughout the background. Earlier in the year the Alpha-Omega project scanned PHP using Anthropic's Claude Mythos, and the team uses OpenAI's Daybreak Red tooling to help check incoming reports. The Foundation says it has published 15 CVEs since 1 July and dealt with more than 100 invalid or duplicate reports. Its summary is worth quoting: AI helps the team validate reports, but a person makes every call.

A messy disclosure

The release did not go to plan. The CVE numbers leaked at about 15:00 UTC on 20 October, so the team brought the release forward to 14:00 UTC on 21 October. A video was published by mistake, and the Foundation says the company Chainguard broke the embargo at 07:41 UTC on 21 October, several hours before the fixes were out. The Foundation has asked hosting companies and businesses that rely on PHP to sponsor its security work.

Claude writing proofs, not just finding bugs

A day later, Anthropic and its partner Theorem published research on a different approach. Instead of hunting for bugs, they asked Claude to produce machine-checked proofs that code behaves safely. A proof in this sense is not an essay. It is a long, formal argument written in the Rocq proof language that a computer checks step by step, so a mistake in the reasoning gets rejected.

Anthropic reports that Claude proved 239 entry points across 123 open-source libraries from Google's OSS-Fuzz project cannot crash on any input smaller than 64 MB. Together the proofs come to about 180 million lines of Rocq, with a typical proof around 600,000 lines. Along the way the work found hundreds of bugs, including rediscovering a known flaw in the libexpat XML library, CVE-2026-56132. The team also proved the Linux kernel's zlib compression code correct against its specification on x86. In a separate effort on a PDF reader, Anthropic says it found as many security bugs in one week as are typically reported in a year.

Anthropic says Claude Opus 5.5 used about 40% fewer tokens than Opus 5 on this work. That is Anthropic's own measurement.

The fine print

Anthropic is careful about what was proved. In its own words, "formally verified does not mean bug-free". The proofs cover single-threaded, statically linked programs compiled for the RISC-V processor design, with a fixed memory budget and a model of how Linux behaves. They prove the code will not crash in ways that a standard memory checker would detect, and that checker does not catch every kind of bug. Not every entry point in every library was covered. This is a strong result for a narrow but important property, not a guarantee that these libraries are now secure.

Why the two stories belong together

The PHP release shows the cost of AI-assisted bug hunting. More reports arrive, many of them duplicates or invalid, and a small security team has to check each one and still ship fixes on time. Anthropic's work hints at a longer-term answer. If AI can produce proofs that a computer can check, defenders can rule out whole classes of bugs instead of patching them one at a time. Neither replaces the other yet, and both still depend on people making the final calls.

What website owners should do

• Check your PHP version. On shared hosting, your control panel usually shows which PHP version each site uses. You need 8.2.35, 8.3.36, 8.4.27, 8.5.12 or later.
• Ask your host. On shared hosting the provider updates PHP itself. If you cannot see the version, ask them to confirm the 21 October fixes are installed.
• Move off old versions. If a site still runs PHP 8.1 or older, plan an upgrade. Those versions no longer receive official releases.
• Static sites are not affected by this flaw. A site made only of HTML files does not run PHP, though other sites on the same account might.