Morning Edition · Tuesday, September 15, 2026Published at 2:27 AM EDT · New York

New research attacks two soft spots in the agent stack: skills and gated memory

SkillAtlas argues skill risk shows up only in execution traces, while BadEngram plants a backdoor in the gated parametric memories that open-weight models use to add capacity cheaply.

Two papers posted to arXiv on Tuesday target layers of the agent stack that most security tooling does not inspect. SkillAtlas addresses agent skills, the reusable units that language-model agents load at runtime. Its argument is that skill…