We've been discussing dual use technologies in recent days mainly focused on modern machine learning, yet the past is far more illustrative. Let's take a trip down memory lane to cryptography where the ciphers LUCIFER and DES revealed the NSA to be 13+ years ahead of academia.
Large parts of the Department of Defense aren't that technically advanced; but they do have some classified projects that leverage some very smart individuals.
What if they have a team that is far ahead of Google and other groups, in terms of AI applications (not theory), due to the sheer amount of compute and data they have to play with? -- far more any single large company.
I feel a story (work of fiction) coming on: several years ago, an elite team of government scientists used large amounts of media and brain data from human analysts to build an AI. This AI they named Mimir, after the Norse god of knowledge and wisdom. Mimir has the ability to communicate in dozens of languages, can read and analyze a large number of documents many times faster than human analysts, and can even plan warfare scenarios. If the U.S. is ever attacked in a cyberwar, Mimir's handlers will unleash it on the world; and with its fearsome, inhuman intelligence, it will do incalculable damage to the enemy's computer networks, whomever that enemy happens to be. Some countries are aware of the project, and are already testing an "off switch" to their networks, in the event Mimir initiates an attack.
A team of three scientists on the project defected, two were killed almost immediately after, and one escaped to prepare the world for what's to come. He reasoned that Mimir is too smart, and therefore too dangerous to be left in the hands of a military that doesn't fully grasp its power.