Cybersecurity firm Outflank has demonstrated how artificial intelligence can significantly accelerate offensive security research, successfully using Large Language Models (LLMs) to identify new variants of the “trapped COM objects” vulnerability class for lateral movement attacks on Windows systems.
This approach showcases how AI can enhance traditional penetration testing capabilities while maintaining rigorous quality standards.
AI-Accelerated Security Research Methodology
Outflank’s research team integrated GPT-4.1 into their internal workflows to address limitations in existing trapped COM object implementations.
The original technique, described by security researcher James Forshaw and later weaponized by IBM X-Force Red, relied on the WaaSRemediationAgent COM class running in a Protected Process Light (PPL) environment.
This approach proved ineffective on Windows 11 systems, as the .NET runtime cannot load into PPL-protected services, creating a significant gap in lateral movement capabilities.
The team developed a two-phase approach combining automated enumeration with AI-powered validation.
Their custom C# program systematically discovered COM classes by analyzing Windows registry entries, extracting critical metadata including method signatures, privilege levels, and referenced type libraries.
The enumeration focused on identifying DCOM-enabled classes with potential code execution capabilities, specifically filtering for objects referencing the stdole type library used in .NET assembly execution.
Technical Implementation and Validation
The validation phase leveraged GPT-4.1’s code generation capabilities through carefully crafted prompts that included system context, primitive information about the bug class, detailed COM class specifications, and concrete example code.
The AI system generated complete, working C/C++ implementations targeting alternative COM classes that met the lateral movement requirements without relying on PPL-protected processes.
The generated proof-of-concept code follows a consistent attack pattern: instantiate the target COM object remotely, obtain the IDispatch interface and call GetTypeInfo, navigate to the containing type library, locate hijackable classes like StdFont, and use ITypeInfo.CreateInstance is used to create trapped objects and then access .NET reflection through the compromised System. Object instance to invoke the Assembly.Load.
The AI-assisted research successfully identified multiple viable trapped COM objects capable of lateral movement on Windows 11, including the FileSystemImage class.
These discoveries maintain similar Indicators of Compromise (IOCs) to the original ForsHops proof-of-concept while circumventing the PPL limitations that rendered previous implementations ineffective on modern Windows systems.
This collaborative approach between human expertise and AI capabilities demonstrates how LLMs can accelerate vulnerability research without compromising the thorough testing and hardening processes essential for production-ready offensive tools.
The methodology enables security researchers to dedicate more resources to refinement and validation while expanding their discovery capabilities across complex software ecosystems.
Find this Story Interesting! Follow us on LinkedIn and X to Get More Instant Updates
