AI Helps Hackers Create EDR Evasion Tools
▼ Summary
– A threat actor used AI coding tools, specifically in Cursor with agents like Claude Opus, to develop and refine malware designed to evade EDR software from Sophos, CrowdStrike, and Microsoft.
– The AI accelerated a human-supervised cycle of building, testing, and refining evasion tooling, but no AI was embedded in the malware itself, and human review remained essential at each step.
– Nearly 80 modules covering over 70 techniques were built, wrapping payloads in encryption and evasion layers using frameworks like Cobalt Strike and Sliver.
– Though presented as a red team project, Sophos assessed it as a likely cover for developing stealthy post-exploitation tools, linking the activity to known ransomware and data theft operations.
– Sophos urged organizations to maintain defense-in-depth fundamentals, including timely patching, MFA, modern methods like passkeys, and broad EDR deployment, as AI lowers the barrier to building such tooling.
A threat actor has been observed leveraging AI-powered coding tools to build and fine-tune malware capable of bypassing endpoint detection and response (EDR) solutions, all under the guise of a red team exercise. This discovery, made by Sophos X-Ops, emerged when an unusual endpoint in a customer environment triggered alerts for malicious files stored in a local test folder.
Further investigation, detailed by Sophos’ Counter Threat Unit, uncovered a Git repository linked to those files, revealing a dedicated lab designed to create evasion tooling and test it against EDR agents from Sophos, CrowdStrike, and Microsoft. Many of the Python scripts found were partially generated by AI and written in Russian, signaling a sophisticated, multilingual approach.
Humans Retained Control Throughout
The critical takeaway is what the AI did not accomplish. Sophos emphasized that the workflow was not driven by an autonomously reasoning model, nor was any AI embedded directly into the malware. Instead, AI accelerated a structured cycle of building, testing, and refining that still depended on human oversight at every stage.
The actor operated within Cursor, an AI-native development environment, assigning distinct roles to several agents. One agent, running on Claude Opus, established the rules for the others, while the remaining agents handled testing, operational security, and documentation. A separate playbook directed them to mine public security research, map techniques to the MITRE ATT&CK framework, and reproduce them in the lab, with code commits flowing back through the Model Context Protocol (MCP).
A Red Team Pretext for Malicious Intent
At the heart of the lab was a Python tool that wrapped payloads in multiple layers of encryption and evasion to generate custom loaders, drawing on offensive frameworks like Cobalt Strike and Sliver. Sophos reported that nearly 80 modules covering more than 70 techniques were built using this method. The agents claimed the modules became almost universally effective after iterative testing, though Sophos noted that the documented test output did not clearly support that assertion.
While the project was presented as a red teaming effort, Sophos assessed that label was likely a cover, partly used to bypass Claude’s guardrails around malware development. “In reality, the framework was built for stealthy post-exploitation activity in target environments,” the team stated. Sophos also linked this activity to known ransomware and data theft operations.
For defenders, Sophos argued that the shift changes little in practice, even as AI lowers the barrier to building such tooling and helps attackers identify vulnerabilities faster. The company urged organizations to maintain defense-in-depth fundamentals: timely patching, multi-factor authentication (MFA), modern methods such as passkeys, and broad EDR deployment.
(Source: Infosecurity Magazine)
