The brief · 13 April 2026

The AI brief, 2026-04-13

Mythos is stepping out of its safe lane to knock on central banks' door. And an open-source model that keeps learning after release changes the question we should be asking about it.

1 min read J / K to navigate

Bessent and Powell Push Banks to Stress-Test Anthropic's Restricted Model, Mythos

The US Treasury and Fed are asking JPMorgan, Goldman Sachs and other lenders to run Mythos against their systems to hunt for security gaps. At the same time, UK regulators are drafting warnings about the cyber risks that same model could pose. Same tool, two opposite readings within days of each other.

practitioners › Mythos, the model Anthropic unveiled in late March as its most powerful yet, is being used to probe for vulnerabilities at JPMorgan and Goldman Sachs even as the UK prepares to flag it as a risk for banks, insurers, and exchanges.

Financial Times →

MiniMax releases M2.7, an open-source model that keeps learning after launch

Most models are frozen the moment they ship. M2.7 promises online learning, meaning capabilities that evolve without a full retrain. Whether that holds up in real-world use or turns out to be mostly marketing remains to be seen.

practitioners › M2.7 is freely downloadable on Hugging Face as of April 13, 2026.

MiniMax →