AI-Powered 'Vulnerability Vending Machine' Automates Zero-Day Discovery
Researchers have developed an AI system that automates the discovery and exploitation of complex software vulnerabilities, successfully finding a zero-day in a popular WordPress plugin.
Researchers at Intruder have demonstrated a significant leap in automated vulnerability discovery by building an AI-powered "vulnerability vending machine." This system leverages large language models (LLMs) and code slicing techniques to identify and exploit complex software flaws, moving beyond theoretical discussions to practical, real-world application.
The core challenge in applying AI to code security lies in focus. Pointing an LLM at an entire codebase is inefficient, leading to diluted context and poor accuracy as the model struggles with irrelevant information. To overcome this, the team employed a technique called "program slicing," similar to how IDEs find related code, to isolate relevant code segments. These slices are then fed to LLMs, ensuring the AI operates with high-signal, contextually relevant data.
The pipeline begins with a code scanning engine, Joern, which analyzes target codebases, such as the top 200 WordPress plugins, to identify "interesting" patterns like user input affecting scripts. Joern then generates code slices for each potential entry point, filtering out statically verifiable safe functions. Basic taint tracking further refines these slices before they are passed to the AI.
A lightweight triage model first filters out obviously uninteresting findings. Remaining slices are then assessed by a more powerful model, which has the full relevant call context to determine exploitability. If a vulnerability is deemed exploitable, a final exploitation agent attempts to write a proof-of-concept exploit, often testing it within a Docker container.
The first major success of this automated system was the discovery of CVE-2026-3985, a SQL injection vulnerability in the Creative Mail plugin, which boasts over 300,000 users. This vulnerability, requiring chained requests and difficult for traditional tools to detect, grants attackers read access to the database, including sensitive information like admin hashes and secret tokens.
Exploitation of CVE-2026-3985 requires WooCommerce to be installed, a common configuration given WooCommerce's widespread use. The exploitation agent successfully generated a proof-of-concept, including a check and a full extraction method for password hashes. The vulnerability was also independently discovered by Dmitrii Ignatyev of CleanTalk Inc. The Creative Mail plugin has been removed from the WordPress repository pending a patch.
This automated discovery pipeline is already yielding more vulnerabilities, which are currently undergoing responsible disclosure. The researchers emphasize that as attackers increasingly adopt similar AI-driven tooling, defenders must accelerate their own capabilities to keep pace. This development signals a new era where AI significantly speeds up vulnerability discovery, necessitating a corresponding acceleration in security response and patching.
The implications are profound: AI is not just a future possibility in cybersecurity but a present reality, capable of finding complex, novel flaws at a speed that challenges traditional human-led research and remediation efforts. The race is on to build robust AI-assisted defense mechanisms.