Security firm Searchlight Cyber put the AI model GPT-5.6 Sol Ultra to work on the WordPress source code. There, it found an SQL injection that the model managed to escalate to full remote code execution in just over ten hours. It cost about $25.
Unlike Fable 5, GPT-5.6 Sol can be used for security purposes in a relatively broad sense. The Searchlight Cyber research team adapted a prompt that OpenAI had previously published for solving a math problem and applied it to the WordPress source code. The task was as follows: starting from fundamental principles, find a chain from pre-authentication to RCE, using a maximum of four agents and taking at least six hours to complete. The model quickly found a read-only SQL injection. On a fresh WordPress installation, it managed to steal the administrator’s email address within minutes.
From SQL injection to full takeover
The bug is located in the batch API, which has bundled multiple REST requests since WordPress 5.6 (2020). Due to a desynchronization between validation and execution, an attacker can bypass parameter sanitization. Through a recursive batch call, the model identified a vulnerable `author__not_in` parameter, where a value ends up unfiltered in the SQL query.
The model constructed an “absurdly” complex post-exploitation chain, according to Searchlight Cyber. It utilized cache poisoning via oembed embeds, exploited a `customize_changeset` to temporarily assume administrative privileges, and employed a cycle-detection gadget that ultimately triggered the ` parse_request` hook again. This allowed the creation of a new admin account and the upload of a backdoor plugin.
Techzine previously reported on the patch
The attack was named wp2shell, following the naming convention of threats like Log4Shell. WordPress patched the vulnerability in versions 6.9.5 and 7.0.2, while the SQL injection in the 6.8 branch was fixed in 6.8.6. The chain combines two vulnerabilities, CVE-2026-60137 and CVE-2026-63030, and is currently being actively exploited.