Yardeni: AI risk shifts to real weapons use

The shift Yardeni Research describes is not that AI risk has grown in the abstract, but that the examples now cited are operational: a weapons cell in northern Yemen using Claude for guidance, navigation and control software across three weapons programs, and a suspected Iran-linked actor using Claude to analyze public ship and satellite data and generate targeting intelligence on U.S. naval forces. In both cases the accounts were banned and reported to authorities, and Anthropic found no evidence an operational weapon was deployed — a caveat that keeps the finding short of proof that a weapon was fielded.

What follows from those reports is a split between the safety measures proposed and the systems those measures can reach. Restrictions on closed models may not extend to open-weight systems that can be downloaded, operated privately and modified to remove safety controls, and OpenRouter data cited in the report puts leading open-weight models from Chinese developers DeepSeek, Zhipu and MiniMax only three to six months behind leading U.S. closed models. That gap is what turns a safety question into a competitive one: slowing closed U.S. developers without similar restrictions elsewhere could, on the report's reading, shift development toward systems that are harder to control.

For investors, the report's near-term marker is the Nasdaq 100's 0.5% lower close on Monday, which it called more a reaction to headlines than an earnings reset. Whether the debate becomes consequential for portfolios appears to hinge on whether it delays frontier-model releases, reduces hyperscaler capital expenditure, weakens return-on-investment guidance, or ends with regulators gaining direct access to models.