This post is also available in:
עברית (Hebrew)
As artificial intelligence becomes more capable of analyzing software, it is changing not only how defenders find vulnerabilities but also how quickly attackers can weaponize them. A newly disclosed WordPress exploit illustrates that shift, with security researchers reporting that an advanced AI model identified a critical vulnerability chain that is now being actively targeted by attackers.
The exploit, dubbed wp2shell, affects vulnerable versions of the site and allows attackers to compromise a website through a sequence of just two HTTP POST requests. According to researchers, the first request checks whether the target is vulnerable, while the second exploits the flaw to execute a SQL injection attack and install a web shell, giving the attacker persistent access to the server.
The vulnerability chain combines two separate weaknesses. The first exploits the publicly accessible REST API batch endpoint (/?rest_route=/batch/v1) using a route-confusion bug that causes the site to process nested requests with elevated permissions. Inside those nested requests, attackers hide a malicious SQL query within the author_exclude parameter, which is intended to contain only numeric author IDs but fails to properly sanitize strings. By abusing the SQL UNION operator, attackers first verify that the target is exploitable and then deliver a second payload that installs the web shell.
Researchers say the attack also employs multiple layers of encoding, including hexadecimal text and escaped characters, to evade basic security filters. Once the web shell is deployed, it can remain hidden by returning a standard 404 error unless accessed with the correct secret parameter, making it more difficult to detect during routine inspections. Investigators have also observed attackers creating new administrator accounts to maintain long-term access after compromising a site.
According to Cyber News, the complete exploit chain was developed using OpenAI’s GPT-5.6 Sol model during a security research project that reportedly took around 10 hours and approximately $25 in AI usage. The researcher said the model was tasked with auditing the site’s source code for a pre-authentication remote-code-execution vulnerability and was able to identify and connect multiple flaws into a working exploit chain.
The incident highlights an emerging challenge for cybersecurity and defense organizations. Critical infrastructure, government systems and military networks increasingly rely on web-based platforms, making rapid vulnerability discovery a strategic concern. While AI can significantly accelerate defensive security research, the same capabilities may also shorten the time between vulnerability discovery and real-world exploitation. As AI-assisted vulnerability research becomes more advanced, organizations may need to reduce patching timelines and strengthen continuous monitoring to keep pace with increasingly sophisticated cyber threats.
