PHP2Shell: AI turned an ignored PHP crash into an exploit threatening WordPress and Drupal sites

A dismissed crash became a web-wide threat

According to IEEE Spectrum, a vulnerability dubbed PHP2Shell came to light in late October and put an enormous slice of the web at risk, including properties run by NASA, Wikipedia and the White House, along with sites built on WordPress and Drupal. The flaw reportedly allowed an attacker to read sensitive data and execute malicious code on the host server.

Disaster appears to have been averted: researchers disclosed the issue to the PHP Foundation, and a patch shipped before criminals got hold of the exploit. But IEEE Spectrum frames the episode as a warning about what happens when AI-assisted security research meets software written across four decades. Matt Suiche, AI security research lead at the startup Tolmo, told the outlet that legacy code written by humans since the early 1990s is now being exposed, and the accumulated technical debt has come due.

PHP dates back to 1994, when the acronym stood for personal home page, and it still underpins a large share of older sites. IEEE Spectrum reports that the bug behind PHP2Shell had been sitting in the PHP Foundation's tracker as far back as 2024. It manifested as an out-of-memory error, where the memory allotted to a piece of data is too small to contain it. Errors of this kind usually just crash the program, which is unpleasant but rarely treated as dangerous, so the report went unaddressed. Suman Jana, an associate professor of computer science at Columbia University, notes that projects always have a backlog of crash-only bug reports that most people do not take seriously.

The audit that changed the stakes

Earlier this year, the security firm Calif offered to independently audit PHP, with sponsorship from OpenAI. Within two weeks, a Calif researcher working with AI tools found a way to escalate the out-of-memory bug into arbitrary code execution on the host server, enough to install malware or read critical data. Calif immediately alerted the PHP Foundation, and the cybersecurity company Wiz later confirmed and expanded the findings, per IEEE Spectrum.

Jana explains that exploiting the bug requires working out how to jump to a specific location in the memory of a particular program, then determining exactly which bits to overwrite and in what order to take control of the system. In his assessment the work is not intellectually exotic, just painstakingly specific reasoning of a kind AI handles well.

He cautions that the humans involved, at Calif, OpenAI and Wiz, had deep expertise in both security and PHP, and that AI clearly cannot do the job alone. Still, it lets the top fraction of a percent of experts accomplish far more, and puts these attacks within reach of people just below that tier who lack command of the low-level details.

Fixing is harder than finding

After disclosure, a team of four engineers spent two months building the patch. Jana points to a fundamental asymmetry: finding a bug needs one working exploit, while fixing it demands closing every avenue of attack without breaking existing behaviour. That burden lands hardest on open-source projects like PHP, which are enormously widely used yet lack clear revenue streams, a chronic problem that AI has made more urgent.

Two developments may help. Frontier labs are donating portions of their token budgets so defenders can probe critical software for bugs, an arrangement Jana would like to see scaled up through coordinated effort or donations. Formal verification tooling is also advancing quickly, including a release from Anthropic on 22 October. In that pipeline, according to Jana, a language model generates the patch and another generates the proof that the patch is correct, with only the proof checker left to traditional software because models hallucinate and the final check must be formal.

Matthew Green of the Johns Hopkins Information Security Institute is more pessimistic. He told IEEE Spectrum that the underlying maintenance problem will go unfunded until something genuinely bad happens, which he speculates could arrive around early 2027. His interim advice is to install security updates the moment they appear rather than deferring them.

Why it matters

PHP2Shell demonstrates that AI has shifted the economics of offence. Bugs long filed away as harmless crashes can now be escalated into full compromise by small teams with model assistance, and the same presumably applies to the vast backlog of legacy code still running the internet. Defence lags structurally, because patching is slower and costlier than finding, especially for underfunded open-source maintainers. Whether the gap closes through formal verification and donated compute, or only after a catastrophic breach, remains an open question. As Jana put it, this will not be the last such vulnerability; the hope is that white-hat researchers keep finding them first.