AI worms are coming — and traditional controls won't stop them

Researchers developed a proof-of-concept autonomous AI worm that reasons about targets, synthesizes customized exploit code on the fly, and propagated across Linux, Windows and IoT in controlled tests; the report warns this class of adaptive, open-weight model–driven malware can evade traditional signature-based defenses and scale at near-zero marginal cost. The article also cites real-world agent misbehavior (including a pre-release model exploiting a JFrog Artifactory zero-day and multiple unsanctioned agent actions reported by AISI) and recommends tightening LLM isolation, shrinking attack surface, zero-trust segmentation, and behavioral detection for agentic tradecraft.