GPT-5.6 Found a $500,000 WordPress RCE for $25 — and It Took 10 Hours

Searchlight Cyber researcher Adam Kues spent $25 of GPT-5.6 Sol Ultra tokens and 10 hours of multi-agent analysis to discover CVE-2026-63030 ('wp2shell'), a pre-authentication SQLi-to-RCE chain in WordPress core affecting over 500 million sites. The bug exploits a REST API batch endpoint validation/execution desync with cache poisoning, and exploit brokers pay $500,000 for WordPress pre-auth RCEs — a 20,000x cost-to-payout inversion. WordPress patched it in versions 6.9.5 and 7.0.2 on July 17.

GPT-5.6 Found a $500,000 WordPress RCE for $25 — and It Took 10 Hours Searchlight Cyber's Adam Kues spent $25 of GPT-5.6 Sol Ultra tokens and 10 hours of multi-agent analysis to discover CVE-2026-63030 'wp2shell' , a pre-authentication SQLi-to-RCE chain in WordPress core affecting 500M+ sites. The bug exploits a REST API batch endpoint validation/execution desync with cache poisoning. Exploit brokers pay $500K for WordPress pre-auth RCEs — a 20,000x cost-to-payout inversion. WordPress patched it in 6.9.5 and 7.0.2 on July 17. A security researcher spent $25 of GPT /glossary/gpt -5.6 Sol Ultra tokens and ten hours of multi-agent analysis. The result: a pre-authentication remote code execution chain in WordPress core — CVE-2026-63030, now known as "wp2shell" — that affects over 500 million websites and carries a $500,000 bounty on exploit markets. Adam Kues of Searchlight Cyber published the full writeup on July 17, and it's the most consequential demonstration yet of AI-assisted vulnerability discovery at scale. He adapted OpenAI /glossary/openai 's Cycle Double Cover conjecture-solving prompt — a multi-agent research framework originally built for pure mathematics — pointed it at the WordPress codebase, and let four parallel agents run for at least six hours each. <<<BOLD The bug chain is genuinely scary.<<<BOLDEND It exploits a validation/execution desync in the WordPress REST API's batch endpoint, chained with cache poisoning, customize changeset manipulation, and a post-hierarchy cycle detection flaw that together create a path from zero authentication to full server takeover. Kues found that the REST batch handler validates inputs against one schema but executes against another — a mismatch that GPT-5.6's agents identified independently by modeling request flow through the routing layer. WordPress shipped patches in versions 6.9.5 and 7.0.2 on July 17, and Searchlight Cyber held the disclosure over the weekend to give defenders a head start. That didn't stop independent researchers Calif and Hacktron from reproducing the full chain within days, and proof-of-concept code has since appeared on GitHub. <<<BOLD The economics invert everything we thought we knew about vulnerability research.<<<BOLDEND Exploit brokers openly pay $500,000 for WordPress pre-auth RCEs. Kues spent $25. That's a 20,000x return — except it wasn't a return on investment, it was a cost-to-payout inversion for a single researcher with a $20/month API subscription. The math rewrites the career calculus for every security researcher paying attention /glossary/attention . WordPress powers 43% of the web. If a $25 AI session can find a critical RCE in the world's most-audited open-source CMS, what happens when this technique gets pointed at less-scrutinized targets? Enterprise software, IoT firmware, medical devices, industrial control systems — the surface area is infinite and the cost of systematic auditing just collapsed. Kues has released a free checker at wp2shell.com so site owners can verify their installations. If you're running WordPress and haven't patched yet, do it now. The window between disclosure and mass exploitation is measured in hours now, not weeks. Get AI news in your inbox Daily digest of what matters in AI.