Channeling Augustus: On Agentic Offensive Information Operations

Channeling Augustus: On Agentic Offensive Information Operations

Photo: New Africa/Adobe Stock

The fixation on defense by commentators thus makes sense: It is of critical importance to understand and counter the content generation capabilities of AI deployed by rivals of the United States. The defenses of the United States must always be better than rival offenses.

The offensive focus on how the United States can best create its own content is also necessary. However, focusing on what to deploy is insufficient. Any good football coach will tell you that (1) the best defense is a good offense and (2) you first must master blocking and tackling.

Particularly in the context of offensive information operations, the targeting and timing of content deployment are as important as the content itself. Put simply, what matters is not just the what, but to whom and when. Here is where AI-powered agents can help.

The Who and the When

As Lindsey Sheppard and I wrote back in 2021, there are real national security risks to the increasing segmentation of the internet. In 2025, the internet is an even more fragmented place. From the Great Firewall in China to similarly active censors in Russia and Iran, penetrating foreign information environments is as challenging as it is critical for offensive information operations.

But it's not impossible. The first step is painting a picture of the target's online ecosystem. What online entities regularly post meaningful content? What platforms do they use? What do they talk about? How influential is their content? Who else is fighting for eyeballs in that information environment?

Existing Tools

Large language models like Claude and ChatGPT, to search engines like Google, Baidu, and Yandex, are extremely useful but, on their own, will not reliably allow you to find, scrape, and analyze high-quality sources of information. Additionally, while more data is accessible for search thanks to AI, this new research tool "makes the problem of sifting through and making sense of such massive quantities of accumulated data qualitatively more difficult." In practice, extracting meaningful signals from raw data streams is both technically complex and resource-intensive.

AI agents can be trained to use the most appropriate language model and relevant search engine(s). They can also figure out precise locations of sources and can regularly evaluate and update source lists. They can work as a team, evaluate one another's work, suggest high-quality sources to human users in a matter of seconds, and be programmed to regularly revisit their work in living, breathing online spaces. The result is a quickly painted, dynamic, transparent, and human-verified portrait of the internet as it exists in the target environment. As CSIS Future Lab colleagues rightly pointed out, "AI agents are only as good as the information they process." Feeding them closely curated-and therefore more trusted-sources in the top-of-the-funnel will produce more useful results than the firehose.

The ability to know to whom content should be targeted should be coupled with knowing when to deploy the content. Here, AI agents can be deployed to scrape content from the sources they found and that humans have validated. Other agents can be used to verify, sort, and filter curated artifacts. Still others can analyze and understand the content, presenting distilled toplines to busy operators and commanders, flagging anomalies or shifts, and even suggesting options for time and place of message deployment. And yes, they can even be trained to generate appropriate content for the appropriate audience to receive at the appropriate time.

The bottom line is that even the best content will be less effective if inappropriately targeted and timed, especially across a fractured internet.

Guardrails

Using agentic information operations, the United States can dive deep into adversary territory for influence. From source curation through to content creation, timing, and deployment, AI agents present useful opportunities. And, as Lindsey and I wrote back in 2021, leaders can use the fracturing of the internet to control their own populations, capabilities since supercharged by AI. The United States must avoid any urges to further fracture the internet and, most importantly, should not deploy these agentic information operations tools domestically.

Beyond our shores, offensive information operations must be conducted in an ethical and morally appropriate way. As a recent piece by uniformed officers in Military Review rightly asserts, "As AI becomes more integrated into military [information operations], ethical considerations must be at the forefront of its deployment." To be clear, the United States does not have to play fair against foreign targets who will not temper their own agentic information operations, but the United States should always look to mitigate the risk to civilians. For example, AI-enabled content should never encourage, condone, or otherwise excuse civilian harm. AI tools can and should be used to mitigate this and other types of risk by simulating thousands of potential outcomes in seconds, helping operators peek around corners and look into the future.

Humans must also have multiple points of control in the loop to ensure standards are upheld. Given our access to top talent, the best and latest technologies that are being developed domestically, thanks to robust entrepreneurial ecosystems, and concerted infrastructure investments, we can still win without stooping to their level.

Augustus may not have worried too much about negative repercussions on innocent civilians, and ancient Rome did not have the internet, so there are limitations to the titular analogy. But he strategically deployed the latest technological tools (e.g., custom gold coins) and undoubtedly would have utilized AI agents for more than just the what, especially in the name of greater peace and stability. And that is a strategy worth channeling.

Erol Yayboke is a senior fellow (non-resident) with the Futures Lab at the Center for Strategic and International Studies in Washington, D.C. He is also the chief operating officer at FilterLabs.AI.

The author would like to thank Jose M. Macias for his valuable contributions.