“World’s Best AI Hacker” Just Uncovered A Massive Web Vulnerability
L ast month, Brian Pak and his researchers at Pentagon contracted startup Xint pointed their AI at NGINX and told it to search for vulnerabilities. NGINX helps run nearly two-thirds of the most visited sites on the internet. A bug there is a bug almost everywhere.
The AI, which Xint describes as “the world’s best AI hacker,” used OpenAI models to find four severe flaws with zero human assistance. Then Pak’s team took the wheel and with some tactical prompts and decades of hacking instinct, pulled out what he describes as the “the holy grail” for hackers of his generation: a vulnerability that lets an attacker run code on an NGINX server remotely, with no authentication required. They could steal data. Run malware to pilfer information like credit card details over time. Or hang around and attack anyone visiting the site.
The issue dates back to 2015 and is the result of NGINX’s mishandling of attacker-supplied data designed to force a server to leak memory creating a path for an intruder to execute commands. F5 Networks, which maintains the NGINX code, confirmed all five flaws found by Xint and said it was patching them Wednesday. F5 is urging people to update now before malicious hackers find and exploit the same bugs. There are as many as 11.8 million vulnerable versions of NGINX currently running on the internet, Pak’s research team claims.
Scores of security researchers have read and analyzed that code for over a decade without turning up the flaws announced today. The AI did it in an afternoon. “Xint went through about 50,000 lines of NGINX code in 90 minutes and came back with four bugs with no human guidance,” Pak, CEO and cofounder at Xint’s parent company, Austin-based Theori, told Forbes . Other cyber researchers are finding much the same. Google-owned cybersecurity business Wiz said it and other companies had separately found the same severe flaws in NGINX over the last month.
It was not an expensive effort either. Finding those five vulnerabilities cost just $1,000 in OpenAI tokens, Pak says, using the GPT 6.1 Sol and GPT-6 Luna models. At that bargain basement price, attackers can be less selective when choosing targets. They can go after a bunch of them at once, skipping the reconnaissance and exploit development that used to slow them down. We’re already seeing evidence of this. Earlier this week, Google reported Russian-speaking hackers running a large language model that continuously searched for vulnerabilities and exploited them, with no humans in the loop.
Xint has company. Earlier this year, Forbes reported on another security startup, the $580 million-valued Depthfirst, whose AI also found a vulnerability in NGINX with the help of a cyber researcher. That one needed a particular calibrated configuration to work, Pak says. The bug Xint found with its human hacker as a guide works on NGINX servers using the default settings, meaning it affects significantly more websites.
Both cases point to the same uncomfortable reality: AI is routinely finding dangerous bugs in software many of us depend on. Earlier this year, OpenAI and partners at cybersecurity companies Calif and Wiz discovered a flaw in PHP , the coding language underpinning the majority of the world’s websites. It was worrisome enough that governments and big companies pulled their sites offline rather than wait for a fix. Earlier this week, Calif disclosed another AI-discovered flaw it found in Java, another widely-used web language.
Pak’s AI agent isn’t particularly specialized; It can spot all manner of flaws across various kinds of software. When Theori announced its DARPA grant earlier this year, it said Pak’s Xint had found vulnerabilities that caused the Android version of encrypted messaging app Signal to crash. It’s also turned up severe issues in Apple technologies, including some at the core of the iPhone operating system that could be used to leak data from device memory. Pak says Apple is working on patches for some of the more severe issues Xint has found. As Forbes reported on Tuesday, OpenAI’s models have become especially adept at finding weaknesses in Apple software . Apple has not yet commented on AI’s ability to expose security loopholes in its products.
AI is getting so good at hacking that earlier this year Pak, who has won some of the most challenging cyber games in the world hung up his competition boots., He just didn’t find them fun anymore. The machines had taken all the fun out of the sport. Now, though, he’s happy to watch as Xint digs up vulnerabilities on “autopilot” or with a bit of human help.