Treasury CIO reportedly racing to get access to Anthropic’s Mythos to hunt for flaws

Overview

It has been reported that U.S. Treasury Chief Information Officer Sam Corcos aims to gain access to Anthropic’s large language model, Mythos, as soon as this week — and that he has directed the Treasury’s cybersecurity team to prepare for AI-driven threats. The move, allegedly pushed by Corcos, would let Treasury engineers probe the model for vulnerabilities and test how it might be abused against financial systems. Fast, decisive, and a little tense. Why the rush? Because the stakes are high.

Why it matters

The Treasury watches over critical financial plumbing — tax systems, sanctions enforcement, market surveillance. A flawed or misused model can open doors to fraud, social-engineering campaigns, or automated evasion of controls. It’s part of a larger trend: regulators and agencies want direct access to frontier AI systems so they can red-team them, not just rely on vendor assurances. That shift from polite requests to knocking on the door, pronto, signals anxiety about how fast capability and risk are evolving.

Tensions and next steps

Access isn’t a simple “here you go.” There are legal, commercial and safety tensions: proprietary code, data privacy, and the risk that probing a model could itself reveal attack vectors. It has been reported that talks with Anthropic and internal preparations are underway; whether disclosure arrangements, scope limits, or on-site testing rules get hammered out this week remains unclear. Expect careful choreography — and some hard questions about transparency versus IP protection.

Watchlist

Keep an eye on whether Anthropic agrees and on any public statements about the scope of testing. If Treasury actually gets in, it could set a template for how government agencies scrutinize high-risk AI — and fast. This isn’t just tech-safety theater. It’s a precautionary sprint to secure money and trust in a world where models can be both the tool and the threat.

Sources: bloomberg.com