Kyle Avery

Training Specialist Models: Automating Malware Development

|

This post complements the presentation I gave at Black Hat USA 2025.

Can a small, self-hosted LLM outperform state-of-the-art models at evasive malware development?
In this technical deep dive, we explore how reinforcement learning with verifiable rewards (RLVR) enables training compact specialist models that rival large generalists in domain-specific tasks.

In the first half of this post, we’ll break down the LLM training process and recent opportunities created by RLVR. The second half details our training methodology for Dante-7B, a 7 billion parameter model that generates functional, evasive Cobalt Strike shellcode loaders capable of bypassing Microsoft Defender for Endpoint (MDE). We’ve released Dante-7B on Hugging Face, complete with a demo app so anyone can experiment with the model.

Tags:

Read full post

Accelerating Offensive R&D with Large Language Models

|

At Outflank, we continually seek ways to accelerate our research and development efforts without compromising quality. In this pursuit, we’ve begun integrating large language models (LLMs) into our internal research workflows. While we’re still exploring the full potential of AI-powered offensive tooling, this post highlights how we’ve already used AI to speed up the delivery of traditional offensive capabilities.

By leveraging AI as a research accelerator, we can dedicate more time to refining, testing, and hardening the techniques that ultimately make it into our OST offering. This post is a case study of our AI-assisted exploration of the “trapped COM object” bug class.

Tags: , ,

Read full post

EDR Internals for macOS and Linux

|

Many public blogs and conference talks have covered Windows telemetry sources like kernel callbacks and ETW, but few mention macOS and Linux equivalents. Although most security professionals may not be surprised by this lack of coverage, one should not overlook these platforms. For example, developers using macOS often have privileged cloud accounts or access to intellectual property like source code. Linux servers may host sensitive databases or customer-facing applications. Defenders must have confidence in their tools for these systems, and attackers must understand how to evade them. This post dives into endpoint security products on macOS and Linux to understand their capabilities and identify weaknesses.

Tags: , ,

Read full post

Unmanaged .NET Patching

|

To execute .NET post-exploitation tools safely, operators may want to modify certain managed functions. For example, some C# tools use the .NET standard library to terminate their process after execution. This may not be an issue for fork&run implementations that spawn a sacrificial process, but executing in-process will terminate an implant. One could write a small .NET program that resolves and patches these functions, but we were interested in an unmanaged approach (i.e. a unmanaged implant executing managed code in-process). While our example targets System.Environment.Exit, a similar technique should work for any managed function.

Tags:

Read full post