A researcher used GPT-5.6 and $25 to find a WordPress zero-day that exploit brokers would pay $500,000 for, by adapting OpenAI's own math-solving prompt

GPT-5.6 Found a $500K WordPress Zero-Day for $25 in Compute

The exact multi-agent prompt that OpenAI published to solve open mathematical conjectures can also find zero-day vulnerabilities. Adam Kues, a security researcher at Searchlight Cyber, proved it by turning that prompt on WordPress source code. Six hours and roughly $25 of compute later, GPT-5.6 had discovered a pre-authentication remote code execution chain in the platform powering over 500 million websites. 1

The cost gap is the story. Exploit brokers value pre-auth RCE bugs in WordPress at roughly $500,000. Kues found one for the price of a takeout order. 1

When GPT-5.6 Sol Ultra launched, Kues noticed that OpenAI had published the prompt it used to solve the Cycle Double Cover conjecture, a longstanding open problem in mathematics. He saw no discussion of applying it to security research. So he took the prompt, which orchestrates multiple AI agents to explore diverse strategies in parallel, and redirected it from mathematical proofs to source-code auditing. 1

The adaptation was minimal. Kues cloned the latest stable WordPress release, stripped out the git history to prevent the model from cheating by diffing code against patched versions, and instructed four agents to spend at least six hours analyzing the code from first principles. The bulk of the prompt was taken almost directly from OpenAI's published methodology: maintain a diverse portfolio of attack approaches, avoid fixating on the most promising lead, keep multiple incompatible research routes alive, and do not stop after the first wave fails. 1

What the agents found was a pre-authentication SQL injection in the WordPress batch API, a feature introduced in version 5.6 in 2020 that lets users bundle multiple API requests into a single call. The endpoint is reachable without authentication. 1

Kues did not believe it at first. WordPress is among the most heavily audited codebases in existence, and it had not seen a meaningful pre-authentication vulnerability in years. To test the claim, he installed a stock WordPress instance on a remote server and asked the model to steal the administrator's email. Within minutes, it printed the address he had used to set up the instance. 1

Then he asked whether the SQL injection could escalate to full remote code execution. About four hours later, the model responded with a chain: the read-only injection could elevate an attacker to administrator without cracking any passwords or performing offline computation. From there, code execution was straightforward. The total cost was 50% of his weekly usage on a $200 subscription, roughly $25. 1

The post-exploitation work, in Kues's description, was "completely absurd." The model wrote it in four hours. It took him far longer to understand. 1

Before disclosure, external researchers identified as Calif and Hacktron independently reproduced the full exploit chain. 1

The same reasoning architecture that lets frontier models solve open math problems can systematically discover zero-day vulnerabilities in production software at commodity compute prices, using a prompt that is published and freely available. For platform operators, the assumption that finding a novel vulnerability requires weeks of expert human labor no longer holds. For the exploit broker market, the supply of high-severity bugs is about to increase. For anyone running WordPress, now is the time to update.