AI exploits a 22-year-old PHP vulnerability in two days

According to a CNN Business analysis by David Goldman (October 23, 2026), a critical vulnerability in PHP, the language that the article says powers 70% of the world's websites, had gone unresolved for 22 years, and artificial intelligence turned it into a practical threat in just two days.

The story dates back to July 2004, when security researcher Stefan Esser warned of a critical flaw in PHP that could let an attacker read passwords, steal databases or even take over entire websites. The warning was largely ignored for decades. In 2020 another researcher discovered a related issue, and PHP's maintainers released a partial fix. In 2024 a similar vulnerability was found and left unpatched.

According to the article, the reason for this inaction was that exploiting the flaw required tedious, slow and monotonous human work, which made it impractical. This is what is described as "security by exhaustion": many vulnerabilities go unexploited not because they are harmless, but because taking advantage of them takes too much effort. AI does not get tired, and that brake disappears.

On July 30, researchers at AI security company Calif, together with OpenAI's "Patch the Planet" security initiative, pointed a large language model at PHP without expecting to find anything critical. On August 1 they found a method to attack PHP's vulnerabilities, and days later they found another, and then another. Over the course of that month, Google-backed security company Wiz, also a Patch the Planet participant, likewise discovered new ways to exploit those vulnerabilities. The patch was released on Wednesday and the companies publicly disclosed their findings; Wiz called it one of the most serious vulnerabilities in the history of software.

Calif researcher Ryan Austin noted in a blog post that no special AI models were used, only models available to the general public. According to the article, Anthropic paused the rollout of its Mythos model in June because it was so effective at exposing vulnerabilities that the U.S. government deemed it a national security threat.

In technical terms, the flaw made it possible to manipulate how PHP processes requests. If a server crashed after running out of memory, it left behind stale data that could be exploited to leak information or execute malicious code. The article compares it to an overwhelmed restaurant kitchen: the chef walks out mid-service leaving ingredients and recipes on the counter, and the next cook serves contaminated food and even sends recipes out to customers. The difficulty lay in forcing the memory failure at a precise moment to bypass PHP's security limits. The AI quickly found two routes: one that exploits the discrepancy between how PHP measures compressed and decompressed files, and another based on deeply nested file structures that overflow PHP's directory structure and trigger the memory failure.

The threat may not be over. Adam Meyers of CrowdStrike warns that many outdated versions of PHP remain in use and that the patch almost certainly does not fix all of its problems. In addition, according to Cloudflare, an apparent leak ahead of the patch triggered thousands of exploitation attempts in the hours before the announcement. Meyers notes that attackers do not need to find the vulnerability first: they can weaponize it after disclosure, and it remains to be seen how much will surface in the wild.

OpenAI's recommendations in its disclosure boil down to writing more secure code from the start, which is of little help to those who already have PHP applications. Migrating away from PHP is expensive, according to Meyers, because it means rebuilding the entire application. Vyas Sekar, a professor at Carnegie Mellon, believes this case could be the first of many and that there are many more vulnerabilities waiting to be discovered and exploited. Nick Rogers (Fieldwork AI) argues that "security by exhaustion" is no longer a viable safeguard, and Madhu Shashanka (Concentric AI) advocates assuming everything is a potential risk, warning that leaving software unpatched, even for a few days, is now far more dangerous because the effort that once protected those systems is no longer a barrier.