'Instruction Bleed': cross-module interference is a security flaw in prompt-composed agents

arXiv·medium signal

arXiv:2606.26356 (June 26) names and analyzes 'Instruction Bleed' — where instructions from one module of a prompt-composed agentic system leak into and corrupt the behavior of another module. This is a structural vulnerability class for the increasingly common pattern of stitching agents together from composable prompt fragments and skills. Builders assembling multi-skill or multi-prompt agents should treat module boundaries as a real isolation problem, not just an organizational convenience.