AI model identifies critical WordPress vulnerability in hours for $25

A security researcher affiliated with Searchlight Cyber has disclosed a critical pre-authentication remote code execution (RCE) vulnerability in WordPress, identifying the flaw using the GPT5.6 Sol Ultra AI model developed by OpenAI. The discovery highlights a significant shift in the landscape of automated security research, as the entire exploit chain was developed in just over 10 hours at a cost of approximately $25 USD in API usage.

The vulnerability stems from a validation desynchronisation bug within the WordPress batch API, which allows an attacker to bypass parameter sanitisation and execute a SQL injection. The AI model identified this initial flaw and subsequently chained it with advanced techniques, including cache poisoning and the abuse of the customize_changeset mechanism, to escalate privileges to an administrator level and achieve full code execution.

To execute the exploit, the model utilised a recursive batch call to bypass restrictions on GET requests, accessing a SQL injection point within the author__not_in parameter of the posts endpoint. The privilege escalation chain involved using the SQL injection to poison the in-memory cache of WP_Post objects and fabricating database rows through the oembed_cache post type. By exploiting a cycle detection mechanism in post hierarchies, the attacker forced WordPress to reconcile conflicting states, allowing control over the post_content field.

The researcher adapted a prompt originally used to solve the Cycle Double Cover mathematical conjecture for this security research application. After identifying the pre-authentication SQL injection, the model determined that the vulnerability could be escalated to RCE without cracking passwords or performing offline computation. The final exploit involved creating a fake customize_changeset to temporarily assume administrator identity and using the parse_request hook to replay the request with elevated privileges, enabling the creation of a new administrator account and subsequent plugin upload.

The vulnerability was responsibly reported to the WordPress project prior to public disclosure to allow for timely patching. The researcher noted that while the AI model produced the exploit chain rapidly, the human researcher spent significantly more time untangling the complex post-exploitation logic. This incident underscores the growing capability of large language models to identify disparate gadgets and chain them across large codebases with precision that rivals or exceeds traditional human security research methods.