Week of November 24, 2025

Security researchers have developed a method using large language models (LLMs) to craft malware that evades behavioral detection by endpoint detection and response (EDR) systems. The technique leverages LLMs to generate code that mimics legitimate user behavior, effectively bypassing behavioral analysis engines. This proof-of-concept highlights a potential new attack vector against modern endpoint security solutions.

Mandiant's M-Trends 2026 report reveals two new malware families, PROMPTFLUX and PROMPTSTEAL, that actively query large language models during execution to dynamically adapt their behavior and evade traditional endpoint detection. PROMPTFLUX uses LLM responses to select evasion techniques based on the target environment, while PROMPTSTEAL focuses on extracting proprietary ML models via distillation attacks. This represents a shift from AI-generated malware to AI-driven runtime adaptation, breaking signature and behavior-based defenses.

A detection rule identifies when a new AMSI provider is added via the Windows Registry, which attackers can use to bypass AMSI protections. This technique allows malicious code to evade script content scanning by registering a rogue AMSI provider. The rule helps defenders detect such persistence mechanisms.

A CrowdStrike insider was terminated after sharing screenshots of internal dashboards, including an Okta SSO panel, with the financially motivated group Scattered Lapsus$ Hunters. The group falsely claimed a system breach, but CrowdStrike confirmed no compromise occurred and customer protections remained intact. The incident highlights insider threat risks to security vendors.

A research proof-of-concept demonstrates using the Nim programming language to evade EDR detection by disabling AMSI and ETW, invoking NT system calls directly, and unhooking ntdll.dll. The article provides a technical walkthrough for educational purposes, highlighting how these techniques can bypass user-mode API monitoring.