Flaws that let attackers run arbitrary programs remotely
Two serious flaws that let attackers run arbitrary programs remotely were found in PHP, a language widely used for programs on website servers. Dubbed "PHP2Shell," the flaws could, if exploited, allow attackers to take over servers, steal confidential data or alter websites. PHP's development team released an emergency fix on Oct. 21 (late at night the same day Japan time).
About 70% of websites whose server-side language can be identified use PHP. The flaws had been in PHP for 14 years and surfaced in a limited form in 2024, but were not considered serious at the time. Finland, the Netherlands and France temporarily took government websites offline on Oct. 21.
Found with OpenAI's model, handled by a coalition over 2 months
OpenAI President Greg Brockman posted on X on Oct. 21 (early Oct. 22 Japan time) that the flaws were found and addressed using the company's model. He said it was part of "Patch the Planet," an initiative OpenAI launched in June to find and fix flaws in critical software with AI, and that the company spent two months forming a coalition with the companies involved.
U.S. security company Calif found the first flaw on Aug. 1 and the second on Aug. 21, and reported them to PHP. It was reported to have been using OpenAI's model. OpenAI's own team also found a more complete form of the flaw in September. U.S. cloud security company Wiz had separately found the same flaw.
OpenAI's Codex Security team, PHP's security team and Google's Gemini team took part in the response. For organizations unable to update right away, Google Cloud, Cloudflare, Amazon, Microsoft and others provided detection methods and temporary mitigations. The fix had been scheduled for release on Oct. 22, but was moved up a day after details leaked early.
| Date | Event |
|---|---|
| Aug. 1 | Calif finds the first flaw |
| Aug. 11 | Calif reports to PHP |
| Aug. 19 | Wiz separately finds the flaw |
| Aug. 21 | Calif finds the second flaw |
| September | OpenAI's team finds a more complete form of the flaw |
| Oct. 21 | PHP releases an emergency fix; 3 European countries take government sites offline |
Defenders faster than attackers
Brockman said that as model capabilities improve, defenders need to move faster than attackers. He also proposed that authorized defenders such as governments search for vulnerable servers and, in extreme cases, patch them remotely. He warned that many more similar issues would be found in the coming months.
Anthropic also launched "OSS Scanner" in October, which uses its top model to examine critical open-source software for free, and competition to find flaws with AI is intensifying.
