Detection Engineering in the Era of Semantic Malware

This article examines the emergence of "promptware"—malicious prompt-injection attacks against tool-using AI agents (e.g., Claude Code, Copilot, Cursor)—highlighting the Brainworm proof-of-concept and a corpus of research demonstrating persistence, C2, lateral movement, and data exfiltration achieved without traditional binaries. It argues that signatures, process-tree, and network indicators often fail for semantic malware and prescribes prevention (least privilege, signed memory files), new agent telemetry standards, context-window and session logging, behavioral baselining, file-integrity monitoring for memory files, and dedicated red-team test cases to enable effective detection and response.