AI Finds the PHP Bug. Attackers Read Along.

Researchers at security company Calif used AI tools and human expertise to develop exploits for vulnerabilities in PHP, the language that powers many websites, Wired reports. After public disclosure, Cloudflare recorded millions of exploitation attempts within hours. Wired describes AI-assisted attacks, but the PHP Foundation said it had seen no successful exploitation. Attempts are not confirmed takeovers.

The flaws, collectively called PHP2Shell, expose a change in what counts as a practical threat. Calif developed exploits in early August, partly through OpenAI’s Patch the Planet program. Some underlying issues had been identified in 2024 but were considered too difficult to exploit. Researchers at OpenAI and Google-owned Wiz independently found PHP2Shell flaws using AI tools. Human knowledge still mattered. AI assistance does not establish autonomous exploitation.

Our earlier PHP coverage examined the gap between releasing a patch and authorizing its installation. Wired’s reporting advances that question: what happens when attackers can use public information before operators complete deployment? The documented sequence runs from research to disclosure to a burst of attempts. It does not supply a complete timetable for protection across affected websites.

Faster discovery can give maintainers a valuable head start when researchers coordinate disclosure. It can also turn a previously impractical weakness into urgent work. Those outcomes are compatible. The attacker does not need approval for a maintenance window. The defender may need testing, a rollback plan, and permission to interrupt a service that takes orders or serves customers.

Hosting arrangements decide where those permissions sit. If a provider manages the PHP runtime, the website owner may depend on the provider’s update schedule. If the customer manages the server, the customer may carry that duty. An application developer may control neither arrangement. These are different operating models, not findings about any particular host in Wired’s account. A dashboard login does not necessarily confer authority to replace the software beneath it.

The reported attack volume makes deployment urgent, but it does not measure successful breaches. Wired notes that many circulating potential exploits appeared ineffective. The Foundation’s report of no observed successful exploitation also has a boundary: what the Foundation has seen. Neither observation justifies announcing a mass compromise or treating every exposed installation as safe.

The next decisions belong to identifiable operators. PHP maintainers can make affected-version and fix guidance clear. Hosting providers can state which customer environments they updated and which still require action. Website administrators can identify who controls their runtime, deploy the applicable update, and verify the running version and service behavior. Providers and site owners can then notify customers about remaining work or disruptions. The useful completion notice describes protection in operation, not merely a patch available somewhere.