OpenAI has hired Dan Driscoll, who served as Secretary of the Army until earlier this year, to look after its national security business, CNBC reported on Nov. 3, the clearest sign yet of how seriously the company is treating the defense and intelligence market.
The company did not disclose his title, and a spokesperson declined to describe the scope of the role beyond saying it covered national security matters. People familiar with the arrangement said he would deal directly with defense and intelligence agencies that are evaluating models for analysis, logistics and software work.
The appointment is the latest in a series of additions to a government-facing business that OpenAI has been assembling quietly. The company has been hiring people with clearance-adjacent backgrounds and building the compliance apparatus that federal buyers demand before they will put a workload on a commercial model. Driscoll’s arrival puts a former service secretary, rather than a sales executive, in front of those customers.
The job he is taking is not the one the title suggests. A Secretary of the Army is a civilian official who runs the service’s budget, personnel and acquisition machinery on behalf of the President, and who spends much of his time arguing with Congress and with the other services about money. That experience is only partly transferable. What it does give him is a map of how the Pentagon buys things, which offices control requirements, and which programs are large enough to support a new vendor.
The harder part is technical. Deploying a commercial model inside a defense network requires accreditation for the environment it runs in, restrictions on where data can be processed, and a supply chain that procurement officers can defend in an audit. Model weights have to be pinned to versions that will not change under a mission, and outputs have to be traceable when a decision is challenged later. A former service chief can open doors; he cannot shorten an accreditation review.
Analysts said the hire reflects a shift in what the defense market wants. For the first several years of the commercial AI boom, the Pentagon’s interest was experimental: pilots, prize challenges, task forces. The buyers now arriving with money are asking for tools that can be fielded, which means vendors need people who understand requirements documents, program offices and the difference between a demonstration and a deployment.
The intelligence agencies bring a different set of demands. Much of their work is reading and summarizing at volume, which is close to what commercial models already do well, but the data is classified and the tolerance for error is low. Contracts in that world tend to be small, sole-sourced and invisible, and they are awarded on the strength of relationships as much as benchmarks.
Competitors have been moving in the same direction. Anthropic and Google have both built public-sector units, and the large defense primes have been signing partnerships with model developers rather than building their own. Palantir has spent two decades becoming the intermediary between Silicon Valley software and military customers, and its valuation reflects how durable that position has been. A model developer that sells directly is choosing to compete with that intermediary rather than go through it.
There is a domestic political dimension as well. OpenAI’s decision to drop its earlier prohibition on military use, made in late 2024, drew criticism from employees and from researchers who had signed on to the company’s safety mission. Each subsequent government contract has been read by that constituency as evidence of where the company is heading. Hiring a former Army secretary will be read the same way.
The revolving door itself is not new. Defense technology firms have recruited senior officers and civilian officials since the first wave of contractors, and the practice is legal within limits set by post-employment rules that restrict former officials from lobbying their old agencies for a period. What is new is that the companies doing the recruiting are now among the most valuable in the world and are hiring for roles that did not exist when the rules were written.
For the agencies, the calculation is simpler. They have budgets for software and a mandate to adopt AI, and they have found that the fastest way to get it is to buy from the companies that make the leading models. The constraint is not willingness but process, and process is what a former service secretary knows how to navigate.
OpenAI has not said how large the government business is or how it is organized inside the company. The hire suggests the answer is large enough to justify a senior executive rather than a business-development team, and that the company expects the next round of defense contracts to be contested on execution rather than on access. Whether a former Army secretary can deliver a model that a program office will put in front of a commander is the question his tenure will answer.
