LLMs Boost Offensive R&D by Identifying and Exploiting Trapped COM Objects

Outflank is pioneering the integration of large language models (LLMs) to expedite research and development workflows while maintaining rigorous quality standards.

This approach allows teams to focus on refining and testing techniques for their Outflank Security Tooling (OST) suite, which delivers evasive capabilities for complex operations.

A recent case study exemplifies this by demonstrating how AI-assisted exploration uncovered viable alternatives within the “trapped COM object” bug class, originally detailed by James Forshaw in early 2025.

Accelerating Research with AI Integration

Forshaw’s analysis highlighted a Windows vulnerability involving trapped Component Object Model (COM) instances, which IBM X-Force Red later weaponized for fileless lateral movement.

Their proof-of-concept (POC) relied on hijacking the StdFont COM registration to point to System.Object, enabling .NET reflection for assembly loading via the WaaSRemediationAgent class over DCOM.

However, this method falters on Windows 11 due to the Protected Process Light (PPL) protections in the WaaSMedicSvc service, which block .NET runtime injection.

Recognizing this as a limitation of the specific COM class rather than the bug class itself, Outflank’s researchers sought non-PPL alternatives that support DCOM and expose code execution primitives, leveraging LLMs to streamline the discovery process.

Outflank’s methodology employs a private framework inspired by multi-agent architectures like Argusee, orchestrating enumeration and validation phases in a unified harness.

During enumeration, a custom C# tool scans the Windows registry for registered COM classes, capturing metadata such as CLSIDs, ProgIDs, server paths, DCOM enablement, interface details including IDispatch support, and referenced type libraries.

This yields JSON outputs filtered for promising candidates, such as those referencing the stdole type library used in Forshaw’s technique for .NET assembly execution via trapped objects.

While automated filtering identifies potential vectors like hijacking StdFont to access System.Object and invoke Assembly.Load it stops short of full automation, routing outputs for human review and further testing to bridge gaps in reliability.

This hybrid model proves efficient, avoiding the complexities of training specialized models for malware development.

Exploitation via Prompt Engineering

Outflank harnessed GPT-4.1 to generate functional C/C++ COM client code for lateral movement POCs, guided by structured prompts that included system instructions emphasizing complete, working code with headers, dependencies, and step-by-step comments.

Prompts supplied primitive information on the trapped object vector (e.g., IDispatch to ITypeInfo navigation, type library resolution, and .NET reflection), registry prerequisites like enabling AllowDCOMReflection and OnlyUseLatestCLR, and detailed COM class metadata.

For instance, targeting the FileSystemImage class (CLSID {2C941FC5-975B-59BE-A960-9A2A262853A5}) with its IFileSystemImage interface and stdole references, the LLM produced code mirroring Forshaw’s ForsHops example, instantiating remote objects, hijacking StdFont, creating trapped System.Object instances, and loading .NET assemblies.

According to the report, This AI-generated code, unedited and shared publicly, enabled lateral movement on Windows 11 by writing to remote files, though it induced client crashes post-execution leaving refinement for production tools to practitioners.

The results uncovered multiple trapped COM objects bypassing PPL restrictions, with indicators of compromise (IOCs) aligning closely to the original POC, including registry modifications and DCOM interactions.

By incorporating concrete examples into prompts, Outflank enhanced LLM reliability for variant discovery, though open-ended research demands adaptive templates.

This AI-driven acceleration not only expands OST’s evasive tooling but also underscores LLMs’ role in offensive R&D, enabling deeper focus on hardening techniques amid evolving Windows defenses.

Organizations interested in OST’s capabilities can explore expert demonstrations to integrate such innovations into their security operations.

Find this News Interesting! Follow us on Google News, LinkedIn, and X to Get Instant Updates!