Pentagon pushing AI companies to expand on classified networks

The Pentagon is pushing the top AI companies including OpenAI and Anthropic ⁠to make their ⁠artificial-intelligence tools available on classified networks without many of the standard restrictions that the companies apply to users. 

During a White House event, Pentagon chief technology officer Emil Michael told tech executives that the military is aiming to make the AI models available on both unclassified and classified domains, according to two ‌people familiar with the matter. 

The Pentagon is “moving to deploy frontier AI capabilities across ‌all ‌classification levels,” an official who requested anonymity told Reuters. 

It is the latest development ‌in ongoing negotiations between the Pentagon and the top generative AI companies over how ⁠the US will use AI on a future battlefield that is already dominated by autonomous drone swarms, robots and cyber attacks.

Michael’s comments are also likely to intensify an already contentious debate over the military's desire to use AI without restrictions and tech companies' ability to set boundaries around how their tools are deployed.

Many ​AI companies are building custom tools for the US military, most of which are available only on unclassified networks typically used for military administration.

Only one AI company - Anthropic – is available in ⁠classified settings through third parties but the government is still bound by the company’s usage policies.

Classified networks are used to handle a wide range of more sensitive work that can include mission-planning or weapons targeting.

Reuters could not determine how or when the Pentagon planned to deploy AI chatbots on classified networks.

Military officials are hoping to leverage AI’s power to synthesise information to help shape decisions.

But while these tools are powerful, they can make mistakes and even make up information that might sound plausible at first glance.

Such mistakes in classified settings could have deadly consequences, AI researchers say. 

AI companies have sought to minimise the downside of their products by building safeguards ​within their models and asking customers to adhere to certain guidelines.

⁠But Pentagon officials have bristled at such restrictions, arguing that they should be able ⁠to deploy commercial AI tools as long as they comply with American law. 

This week, OpenAI reached a deal with the Pentagon so that the ​military could use its tools, including ChatGPT, on an unclassified network, which has been rolled out to more than ‌3 million US Defense Department ⁠employees.

As part of the deal, OpenAI agreed to remove many of its typical user restrictions although some guardrails remain. 

Alphabet's Google and xAI have previously struck similar deals. 

In a statement, OpenAI said this week's agreement is specific to unclassified use through genai.mil.

Expanding on that agreement would ‌require a new or modified agreement, a spokesperson said.

Similar discussions between OpenAI rival Anthropic and the Pentagon have been significantly more contentious, Reuters previously reported.  

Anthropic executives have told military officials that they do not want their technology used to target weapons autonomously and conduct US domestic surveillance.

Anthropic's products include a chatbot called Claude. 

“Anthropic is committed ​to protecting America's lead in AI and helping the US government counter foreign threats by giving our warfighters access to the most advanced AI capabilities,” an Anthropic spokesperson said.

“Claude is already extensively used for national security missions by the US government and we are in productive discussions ‌with the [US] Department of ⁠War about ways to continue that work.”

US President Donald ​Trump has ordered the Department of Defense to rename itself the Department of War, a change that will require action by Congress.