Newly shared research from Tech Against Terrorism shows that three in five AI models failed their terrorism safety test.