IEEE Spectrum reports Calif researchers used AI assistance to turn a long-ignored PHP out-of-memory crash into remote code execution. The flaw, now called PHP2Shell, allows unauthenticated attackers to run code or read process memory. It affects any PHP server, including WordPress and Drupal sites like NASA and Wikipedia, and a patch shipped before known exploitation.
The obvious reading is that AI has made bug hunting cheap and defenders are finally catching up. That reading is premature.
Calif's own claim, as relayed by IEEE Spectrum, is narrower. An expert team that already understood PHP internals found the exploit path within two weeks of an OpenAI-sponsored audit, after the same crash report sat ignored since 2024 because it only caused crashes. Suman Jana, associate professor of computer science at Columbia University who studies automated vulnerability discovery, said the exploit requires painful, specific reasoning to overwrite exact memory bits, work AI now does well for skilled operators, not alone. I find that distinction decisive, because it means the tool amplifies the top 0.1 percent rather than replacing them.
Finding remains faster than fixing.
Once Calif disclosed the chain, four PHP Foundation engineers spent two months building a patch that closes all attack paths without breaking existing functionality, per IEEE Spectrum. That asymmetry is structural. Anthropic has demoed LLM-generated patches paired with LLM-generated proofs checked by a formal verifier, but Jana notes the checker itself must be non-LLM because models hallucinate, and the pipeline is not yet routine for legacy C codebases like PHP. Matthew Green, associate professor at the Johns Hopkins University Information Security Institute, is blunter and says systemic funding will not arrive until a breach angers normal users.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
What PHP2Shell leaves out matters for startups. The report says disaster was averted with no known exploitation, but it offers no telemetry on how many internet-facing PHP installs remain unpatched today and no evidence that similar crash bugs are not already being probed elsewhere. Token donations from frontier labs to audit legacy code, which Jana praised, are discretionary, not a durable funding model for open source maintenance that Matt Suiche, AI security research lead at Tolmo, says carries 40 years of technical debt.
The Qualys advisory not detailed in the IEEE Spectrum piece frames the scale. PHP2Shell maps to CVE-2026-109618 and CVE-2026-109619, each scored CVSS 4.0 9.2 for unauthenticated remote code execution or memory disclosure, per Qualys.
To be fair, the strongest signal for defenders is real. AI moved a bug class humans routinely deprioritized as mere crashes into a working remote exploit quickly, which suggests many similar old reports could be triaged faster now if model access is broadened to maintainers.
If you run PHP anywhere in your stack, patch now and build as if the next crash report you ignore is already weaponized.
