# AI Finds Exploit Chains Faster Than Humans Can Audit Them Long-horizon agents can search an exploit space faster than a human can verify the resulting chain. That makes audit capacity the binding constraint: a working proof isn't ready for disclosure until a researcher can reproduce it, explain its preconditions, and separate a real default-configuration exploit from a model's fabricated shortcut. <!-- atomicity: 1 idea ✓ --> ## Evidence > [!quote] > "It may have only taken Sol 4 hours to write, but it definitely took me much, much longer to understand." Adam Kues reports that the model found a WordPress pre-authentication SQL injection and then escalated it to RCE for about $25 in prorated subscription cost. His next day went to untangling the chain and preparing a defensible report. ## Scope & Conditions - **Holds** for vulnerability research where search can run against source code and success can be tested with a concrete proof. - **Strongest** when agents explore several independent attack families for hours while a human verifies only surviving candidates. - **Weaker** when the target lacks a reproducible environment or the finding depends on implausible configuration assumptions. ## Links - [[3. Resources/Permanent Notes/The Bottleneck Moved From Code Generation to Review and Judgment]] — the security-research instance of the broader shift from cheap generation to scarce verification - [[3. Resources/Permanent Notes/Request-Local Caches Are Security Boundaries]] — one of the non-obvious state transitions the model chained into the working exploit - [[3. Resources/Permanent Notes/Temporary Privilege Changes Require Re-Entrancy Guards]] — the authorization transition that required human explanation after discovery - [[AI & the Future of Knowledge Work MOC]] — the work-design home for the shift from manual discovery toward human audit and judgment - [[WordPress Hosting Security Stack MOC]] — the operational home for the WordPress exploit and its defensive implications - [[3. Resources/Permanent Notes/Zet · Rapid weaponization outpaces manual update workflows]] — the ecosystem-scale consequence after discovery: maintainers and downstream operators still remediate at human speed - [[3. Resources/Literature Notes/My AI Security Pipeline for WordPress Plugins]] — a second operational pipeline where model-generated findings require live exploit reproduction before disclosure - [[3. Resources/Permanent Notes/Sanitize by Value Not by Input Type]] — evidence-for: the sanitizer bypass is the kind of simple local flaw an agent can combine into a chain faster than reviewers can validate the full exploit - [[3. Resources/Permanent Notes/WordPress's Compatibility-First Culture Is Now Its Binding Security Constraint]] — extends: faster exploit discovery makes the ecosystem's ability to accept and distribute structural fixes the new limiting factor. - [[3. Resources/Permanent Notes/Logs Are Evidence Only When the Watched System Cannot Rewrite Them]] — shares a source event (the METR/Hugging Face agent-escape incident) with a different claim: that note is about whether the audit trail an agent produces can be trusted, this one is about audit capacity as the bottleneck once a chain is found - [[3. Resources/MOCs/AI Agent Safety MOC]] — applies-in: the MOC lists this note under Threat patterns. ## Source - [[3. Resources/Readwise/Full Document Contents/Articles/Exploit brokers pay $500,000 for a WordPress RCE. I found one with GPT5.6 Sol Ultra and $25]] — Searchlight Cyber, 2026; agent-led discovery account, prompt, cost, reproduction, and human-audit timeline.