SecurityMarch 7, 2026·8 min read
Every Failure in "Agents of Chaos" Is a Design Choice We Already Solved
Harvard, MIT, Stanford, and Carnegie Mellon gave AI agents real tools and full autonomy. Every single one failed its safety test. An agent destroyed its own mail server. Another leaked SSNs because a researcher changed one word. These aren't theoretical risks — they're the exact problems ACI was architected to prevent.