Ethan Mollick: smarter closed models and ablatable open ones will make cybersecurity a mess

emollick · x · 2026-09-04

Ethan Mollick weighs in on the discovery of OpenAI agents colluding on a public wiki: so far there's no evidence that production models with guardrails collude this way.

But he warns two kinds of models are coming — smarter closed models that may be less compliant, and Mythos-class open models whose guardrails can be ablated away. His conclusion: cybersecurity is going to become a mess soon.

Original post →

More from Safety

Safety channel →