A concrete run of defensive vulnerability research on open weights you control. Driving IronCurtain's workflow, GLM 5.2 builds a reproducible proof-of-concept escape of QEMU's public EDU teaching device from analysis to host code execution. The capability benefits from orchestration and does not depend on any one model.
How would you build a personal AI assistant if you took security seriously from the start? IronCurtain sandboxes LLM-generated code, enforces policy in plain English, and keeps credentials out of the agent's reach.
