OpenAI has identified two critical security flaws in the PHP programming language using its GPT-5.6 Sol model. The vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619, allow unauthenticated attackers to corrupt memory or reuse it during standard web requests. This could let a malicious actor run their own code on a server, a risk that OpenAI said it detected before public disclosure.
AI Finds Flaws in a Core Web Technology
The bugs affect how PHP handles memory and error messages. An attacker needs specific conditions to trigger the memory corruption, but OpenAI demonstrated the attack on Apache servers with default settings. The model found the root cause independently, though other security firms also reported similar issues around the same time.
Calif.io and Wiz flagged overlapping problems in August. OpenAI submitted its report on September 18. Patches were released on October 21 for PHP versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. The Foundation pushed the release date up by one day because CVE identifiers began circulating publicly on October 20, forcing an emergency schedule shift.
OpenAI framed the discovery as evidence that frontier AI models can spot security holes before hackers exploit them. The company expects similar findings to increase sharply in the near future. The model identified the flaws by analyzing the language’s memory and error-handling code, which runs on nearly every web request.
Volunteer teams now face a surge in vulnerability reports. The PHP Foundation says AI-assisted research is raising the volume of submissions. Since July, the security team has published 15 CVEs and dealt with more than 100 invalid, duplicated, or unreproducible reports. Each finding still requires a human to reproduce the issue, assess its impact, and design a fix that does not break existing software.
Fixing the PHP2Shell vulnerabilities proved complex. The team targeted low-level code, balancing backward compatibility, denial-of-service risks, and third-party extensions. The Foundation says it could not have completed the work without dedicated funding and developers working through weekends.
The economic shift is already visible. Rodion Sorokin, Co-Founder and CTO of The Capital Index, noted that finding bugs has become cheap while fixing them remains expensive. He pointed to the PHP Foundation’s $784,376 budget last year, which funded 11 developers. Maintainer capacity, however, has not grown to match the volume of AI-generated findings.
Coordination efforts also struggle under the new pace. Reports from different AI tools often describe different symptoms of the same underlying flaw. This makes it harder to group findings and coordinate a response. If AI systems surface serious flaws at a much higher rate, maintainers will spend less time hunting bugs and more time triaging and validating reports under pressure. A small project run by volunteers cannot absorb the extra workload, even though its code may sit inside thousands of production systems.
Disclosing issues is becoming more difficult as well. The premature publication of PHP2Shell fixes after CVE numbers were already public shows how rapidly maintainers can be pressured when vulnerability details circulate before a coordinated release finishes. Reports produced by AI should supply reproducible proof and, when possible, regression test suites so that maintainers can confirm the problem without starting from zero. Companies that build AI discovery platforms ought also to contribute resources for triage, not merely for finding bugs.
The Cost of Discovery Outpaces Maintenance Capacity
AI-driven bug hunting is reshaping the cost structure of open-source security. The limiting factor is no longer just locating defects; it is confirming them. While AI can highlight many potential issues at a volume beyond manual inspection, each discovery still requires checks for reproducibility, impact evaluation, remediation planning and regression testing before it can be published. For maintainers who volunteer their time, this often translates into additional unpaid effort, especially when AI-generated alerts include false alarms or duplicate reports.
Redefining Security in an AI Era
“AI has made finding bugs cheap. Fixing them is still expensive, and the bill lands on volunteers.” This statement captures the cost shift. “OpenAI expects similar findings to increase sharply.” Nonetheless, the ability of maintainers to handle the influx has not grown proportionally.
