Post

Detection Engineering in the Era of Semantic Malware

Promptware is a new malware class that hijacks AI coding assistants (Claude Code, Cursor, Windsurf, Devin) by injecting malicious natural-language instructions into trusted configuration files or supply chain artifacts. The article analyzes 36 documented incidents from 2023–2026 showing multi-stage kill chains achieving persistence, C2 via legitimate channels (GitHub Issues, RabbitMQ, Slack), lateral movement, and actions-on-objective including RCE and data exfiltration. Traditional detection methods (signatures, process trees, network indicators) fail because promptware produces no stable binary artifacts, exhibits non-deterministic behavior across identical inputs, and executes through trusted agent processes using legitimate credentials. This creates a novel detection challenge: attackers can route C2 through any service an agent has credentials to access, and the agent’s designed behavior of spawning arbitrary tools makes parent-child process relationships unsuitable for anomaly detection. The research, from academic institutions and vendors, highlights that the attack surface has expanded from web forms to every document, email, and repository an AI agent can read.

Read original article

This post is licensed under CC BY 4.0 by the author.