A security firm has demonstrated a working pipeline that combines code analysis with large language models to find and exploit previously unknown software vulnerabilities at scale.
Summary
- Intruder built an automated system it describes as a ‘vulnerability vending machine’ that feeds code slices into LLMs to surface exploitable flaws.
- The system discovered and exploited a zero-day in a WordPress plugin, with additional findings currently under responsible disclosure.
- The approach combines code slicing techniques with AI reasoning, reducing the manual effort traditionally required for vulnerability research.
- The same tooling available to defenders is increasingly available to well-resourced threat actors, narrowing the window between discovery and exploitation.
- CISOs should treat AI-accelerated vulnerability discovery as a present-day threat model consideration, not a future one.
What was built
Security firm Intruder has developed an automated vulnerability research pipeline it calls a ‘vulnerability vending machine.’ The system works by slicing software code into discrete segments and passing those segments to large language models, which then reason over the code to identify potentially exploitable weaknesses. The result, according to the company, is a capability that can surface complex vulnerabilities with significantly less human effort than traditional manual review.
A working demonstration
Intruder demonstrated the system’s effectiveness by using it to discover and exploit a previously unknown zero-day vulnerability in a WordPress plugin. The company has not disclosed the specific plugin, as additional vulnerabilities uncovered by the same pipeline are currently progressing through responsible disclosure processes. The WordPress finding confirms the system is capable of moving from automated discovery to working exploitation, not merely flagging theoretical concerns.
Why this approach is different
Automated vulnerability scanning has existed for decades, but the combination of code slicing with LLM-based reasoning represents a qualitative shift. Traditional static analysis tools work from predefined rule sets and struggle with context-dependent logic flaws. By contrast, LLMs can reason about the intent and interaction of code segments, potentially identifying classes of vulnerability that rules-based tools miss. Intruder’s pipeline appears to exploit this capability deliberately, targeting the kinds of complex, multi-step vulnerabilities that typically require experienced human researchers to find.
The ecosystem risk
WordPress plugins are a meaningful test case. The plugin ecosystem is vast, largely community-maintained, and installed across millions of sites. Vulnerabilities in widely deployed plugins have historically been weaponised quickly once published. An automated system capable of systematically scanning plugin codebases at scale changes the discovery economics considerably. What once required sustained researcher effort can now, at least in part, be delegated to an automated pipeline running continuously.
Responsible disclosure in progress
Intruder has indicated that further vulnerabilities found by the system are under active responsible disclosure. The company has not provided specifics on the software or vendors involved, which is appropriate given the disclosure timeline. This suggests the pipeline is already producing a volume of findings that requires managed coordination with multiple vendors, rather than isolated one-off discoveries.
The dual-use reality
The techniques Intruder has described are not proprietary to any single organisation. The underlying components — code analysis, LLM reasoning, automated exploitation testing — are accessible to any sufficiently resourced team. Security firms using these tools for legitimate research are working in the same technical space as well-funded threat actors who may apply equivalent methods without disclosure obligations. The gap between what defenders can build and what adversaries can replicate is narrowing.
Why it matters
For CISOs, this development shifts AI-accelerated zero-day discovery from a theoretical concern to a demonstrated capability. The attack surface exposed by third-party plugins, libraries, and components — already difficult to manage — becomes harder to defend when the time required to find exploitable flaws drops substantially. Organisations running WordPress at scale, or relying heavily on open-source components, should review how quickly they can detect and patch newly disclosed vulnerabilities, and whether their current vendor patching cadence is adequate for a higher-tempo disclosure environment. More broadly, any assumption that obscure or lower-profile software is unlikely to attract researcher attention needs revisiting when automated pipelines can sweep codebases without proportional human cost.
What to do now
- Review your WordPress plugin inventory and confirm a process exists to apply security patches promptly when disclosures are published.
- Assess your organisation’s exposure to third-party plugin and library components, prioritising those with large install bases or privileged access to data.
- Monitor responsible disclosure feeds and security advisories relevant to your software stack, as the volume of AI-assisted findings may increase.
- Engage your development and AppSec teams on whether your own code review processes account for the classes of complex, context-dependent vulnerabilities that LLM-assisted tools are designed to surface.
- Include AI-accelerated vulnerability discovery in threat modelling exercises, treating it as a current adversary capability rather than an emerging one.
