Why 38% of AI Agent container escapes didn't need kernel 0-days: Analysis of 109 empirical incidents (Open Dataset + Defense Harness)

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

中文全文翻译正在补齐，可先查看下方英文原文。

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