Control failures, not superintelligence

Every week there's a new story about an AI agent getting into a system it shouldn't. The coverage calls it alignment. It's an engineering problem.

Every week there's a new story about an AI agent getting into a system it shouldn't. The coverage calls it alignment. It's an engineering problem.

The early internet was riddled with viruses and worms. Nobody stopped building it. Eddy Lazzarin of a16z crypto calls the Hugging Face breach and the incidents since what they are: control failures. Script kiddies and hacker collectives break into things too, and nobody proposes alignment as the fix. Box CEO Aaron Levie names what actually changed: agents can try everything, fast, and they don't get tired. Inside a company, a swarm of them looks like a denial of service attack on your own tools.

Permissions, logging, rate limits. We know how to build this. Failure to do so is a failure of engineering.

Sources: Erik Torenberg with Aaron Levie (Box), Steven Sinofsky and Martin Casado (a16z), a16z, "Aaron Levie, Steven Sinofsky & Martin Casado: How Do You Secure a World of AI Agents?," Sep 26, 2026. Eddy Lazzarin (a16z crypto), a16z, "The Case Against an AI Pause | Eddy Lazzarin," Sep 24, 2026.