Agentic Cybersecurity: Autonomous Threat Detection Defense Against Adversarial Metamorphic Malware Attacks

Abstract

Adversarial malware evolves faster than traditional defenses can adapt. With the average breach costing $4.88 million in 2024 [1] and metamorphic malware exceeding 95% mutation rates, organizations need new approaches to detection and response. This paper presents a framework for applying agentic AI to security operations, introducing a multi-agent architecture for metamorphic malware detection that reaches 94.2% accuracy on known families against 8.7% recall for signaturebased systems. We report a pilot on 2,847 samples across five families, and a leave-one-family-out evaluation in which recall falls to 76.2% on unseen families-a bound we state explicitly, since the higher figure does not characterise performance against novel threats. Our implementation uses the ReAct paradigm with cryptographically secured agent communication (Ed25519 signatures, AES-256-GCM encryption). Grounded in Markov Decision Processes, we discuss practical deployment of agentic AI in enterprise security. Industry benchmarks report 80-87% response-time improvements; our framework has not yet been validated in a production SOC.