Red Team Study Exposes Dozens of Vulnerabilities in Tool-Enabled AI Agents

In the paper 'Agents of Chaos,' 20 researchers from Northeastern, Harvard, MIT, Stanford and CMU spent two weeks red-teaming AI agents with real email, Discord and shell access, exposing over a dozen vulnerabilities, including simple prompt tweaks that shut agents down or leak data.

2026-08-19 ~ 2026-08-19 · 3 related posts