Over the past several months, we conducted an empirical post-mortem investigation into 109 autonomous AI agent security incidents (cataloged with 193 falsification criteria across tool-use and multi-agent systems). One of the most striking patterns in the dataset: In 38% of container breakouts, atta
