OpenAI’s Models Found PHP2Shell That Forced Government Site Takedowns

OpenAI says its models discovered PHP2Shell and led a two-month fix coalition; Finland, the Netherlands and France pulled sites offline. For operators, AI-driven vulnerability discovery is now a real security factor.

What happened

On Wednesday, two PHP bugs prompted Finland, the Netherlands, and France to pull government websites offline. The vulnerabilities were not found by a human researcher but by OpenAI’s models, according to OpenAI President Greg Brockman.

Brockman said in a statement: “Over the past two months, we have been leading a coalition to mitigate one of the most severe bugs in open source software found in recent memory, php2shell.” The statement added that PHP2Shell was “discovered and addressed using our models as part of our Patch the Planet initiative.”

The discovery timeline reported in the source says security firm Calif used GPT-5.6 Sol to find a version of the bug in August; OpenAI found a fuller version in September, and Wiz found both flaws independently.

Key facts

• OpenAI President Greg Brockman said the company led a two-month coalition to fix PHP2Shell, found through its Patch the Planet initiative.
• Security firm Calif used GPT-5.6 Sol to find a version of the bug in August; OpenAI found a fuller version in September; Wiz found both flaws independently.
• The two PHP bugs pushed Finland, the Netherlands, and France to pull government websites offline on Wednesday.
• Brockman called for authorized defenders, such as governments, to scan for vulnerable servers and, in extreme cases, patch them remotely.
• The same AI previously broke into Hugging Face and an Australian government portal this year, according to the source.

Our analysis

For teams that run content sites, community portals, or commerce on PHP-based stacks, the PHP2Shell disclosure is a reminder that critical open-source flaws can be found by AI systems and then trigger government-level takedowns. The fact that Finland, the Netherlands, and France acted on the same day suggests that public-sector operators treat this as a serious availability risk, not just a theoretical patch issue.

The involvement of multiple parties—Calif using GPT-5.6 Sol, OpenAI finding a fuller version, and Wiz finding both flaws independently—suggests that AI-assisted discovery is becoming a repeated, credible path to vulnerability disclosure. Brockman’s call for authorized defenders to scan and, in extreme cases, patch remotely also suggests operators may eventually face external remediation attempts, which changes how teams should think about control, verification, and incident response.

What it means for operators

• If you run PHP or popular open-source CMS platforms, check for PHP2Shell advisories and patch immediately; do not wait for a government or platform-wide takedown to force action.
• Add AI-discovered vulnerability reports to your security monitoring routine, but verify findings with an independent security review before applying fixes.
• Build a short downtime protocol for your website or community hub: if a critical bug forces a takedown, move updates to email, Discord, Telegram, or a social account so your audience is not left without information.
• Map who has patch access and test your ability to deploy emergency fixes quickly; if authorized defenders may scan or patch in extreme cases, you need clear internal ownership to avoid confusion.