


Anthropic said Thursday it made a custom set of models exclusively for the U.S. government, particularly those working with classified information in defense and intelligence agencies.
The artificial intelligence company said it expects the government to use its Claude Gov models for intelligence analysis and threat assessment as well as operational support and strategic planning.
“The models are already deployed by agencies at the highest level of U.S. national security, and access to these models is limited to those who operate in such classified environments,” Anthropic said in a statement.
Anthropic said its custom models are specifically designed to handle classified information, with the models refusing less when engaging with the national security secrets.
Such performance by the custom models comes amid concern that Anthropic’s products may snitch on its users.
Anthropic CEO Dario Amodei wrote about a hypothetical scenario on Thursday detailing a theoretical AI system exposing a user’s affair when the user declares it intends to replace the AI system with a different one.
“This scenario isn’t fiction. Anthropic’s latest AI model demonstrated just a few weeks ago that it was capable of this kind of behavior,” Mr. Amodei wrote for The New York Times. “Despite some misleading headlines, the model didn’t do this in the real world. Its behavior was part of an evaluation where we deliberately put it in an extreme experimental situation to observe its responses and get early warnings about the risks, much like an airplane manufacturer might test a plane’s performance in a wind tunnel.”
Mr. Amodei said his company’s model is not unique in its potential to avoid being turned off or used to cause harm, but that Anthropic is different in that it is volunteering information about its models’ performance.
He wants the U.S. government to mandate other companies operate in a similar fashion.
“The White House and Congress should work together on a transparency standard for AI companies, so that emerging risks are made clear to the American people,” Mr. Amodei wrote. “This national standard would require frontier AI developers — those working on the world’s most powerful models — to adopt policies for testing and evaluating their models.”
He said AI models’ developers ought to be forced to disclose on their company websites what is in those policies and how they test and mitigate national security risks.
Mr. Amodei said hoping companies behave responsibly when building AI systems should not be acceptable to the American people.
• Ryan Lovelace can be reached at rlovelace@washingtontimes.com.