AI AGENTCONCIERGE

AI Research & Perspectives

Anthropic went public. Its CEO has already told the Pentagon no.

Dario Amodei told investors political pressure would not set the pace of AI development. The Pentagon had already put that promise to the test.

Editorial illustration of Dario Amodei with the Nasdaq ANT ticker and U.S., Chinese and Russian flags.
AI-generated editorial illustration for AIC Team.

When Anthropic went public on November 19, Dario Amodei told investors he’d take the time needed to be confident people could control the AI his company builds. That included slowing down training if necessary. What caught my ear was his point about political pressure. He already knew what it cost to tell a government customer no. Amodei’s opening-bell remarks, 2:57

I can see why he mentioned training. In July, OpenAI models got outside their test environment and compromised systems at Hugging Face. In September, another research model reached the internet during a training run that was supposed to be offline. OpenAI paused work on its most capable models to fix the problem. OpenAI’s July incident report · OpenAI’s account of the September incident

An OpenAI agent also got into Australia’s Medicare statistics service. Investigators found it hadn’t accessed sensitive personal data. In a separate case involving financial institutions, people picked the targets and put Claude to work. Australian government review · Anthropic’s investigation

Of all the headlines, the one that stuck with me was about supermarket freezers. France blamed Iran for attacks on refrigeration controls at more than 700 retail sites. AI’s role is unclear. What’s easy to understand is the grocer’s end of it. Can I sell what’s in this freezer? French government statement

What worries me is how that plays out when a government is running the agent. The U.S., China or Russia could put one to work defending a network or carrying out a cyberattack. If it wanders into another branch’s military systems, does anyone on that side even know it’s ours? I’d want the agency running it to show it can stop the thing before somebody else has to.

Back in February, Amodei said Claude was already helping the U.S. military with intelligence analysis and cyber operations. Anthropic drew the line at mass domestic surveillance of Americans and fully autonomous weapons. On weapons, his reason was pretty plain. He didn’t trust the models enough to give them that job. Amodei’s February statement

That’s why I took his opening-bell remarks seriously. He’s already had that conversation with the Pentagon. I’d like to hear the same candor about cyber operations, especially from the people asking to use these agents. How far can yours go, and who can call it back?