RS-2026-107 — T317-SHAPESHIFTER: Reinforcement Learning-Driven Payload Mutation Achieving 98.8% EDR Evasion in Controlled Adversarial Testing
Year: 2026
Type: preprint
Abstract: RS-2026-107 — T317-SHAPESHIFTER This research presents T317-SHAPESHIFTER, a reinforcement learning-driven payload mutation engine developed within Red Specter Security Research Ltd's NIGHTFALL offensive security framework. Evaluated in the controlled SPECTER BATTLE LAB environment, the system recorded 395 non-detections across 400 test submissions, achieving a 98.8% evasion rate against eight containerised EDR behavioural simulations. The research examines adaptive payload mutation, detection feedback, convergence limitations and the role of cryptographically attested evidence in adversarial security testing. Experimental evidence is described as protected using SHA-256 Merkle commitments, Ed25519 and ML-DSA-65 signatures, and RFC 3161 timestamps. Research limitation: All EDR targets were behavioural simulations operated within an isolated laboratory. The findings do not establish evasion performance against commercial EDR products. Published by Red Specter Security Research Ltd, Chester, United Kingdom. Research reference: RS-2026-107
Author: Red Specter Security Research
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236836924
