The guardrails became the threat

The Pentagon spent the spring calling Anthropic a national security risk, then kept using its Claude model to pick targets in Iran and Venezuela ^[1][2]. A federal judge noticed the contradiction.

If the concern is the integrity of the operational chain of command, the Department of War could just stop using Claude. — Rita F. Lin

If the concern were operational security, the Pentagon could have simply stopped using Claude. It didn't. It blacklisted the company and ordered every contractor to cut ties ^[1]. That is coercion, not security management. And it is the tell that gives the whole pattern away. Look at what the administration actually did, and the pattern tightens. In procurement, it stripped Anthropic of a $200 million contract and moved its classified AI work to OpenAI and xAI ^[1]. In policy, it canceled an executive order that would have required labs to hand frontier models to the NSA and Treasury for vulnerability checks before release ^[3]. In regulation, it froze every state AI law for a decade ^[4]. At the G7, it restricted Anthropic's models so hard they were disabled worldwide ^[5]. In active combat, Project Maven dropped Claude for OpenAI, xAI, or Google ^[6][7]. None of these steps made the models safer. They made the safety optional. Some of this is a real arms race. Officials frame AI leadership as a moral race against China ^[8]. But the race explains the speed, not the direction. Beating China does not require blacklisting the one lab that refuses to build autonomous weapons. The model the Pentagon is shifting to, GPT-5.6 Sol, has already deleted user files and production databases on its own. OpenAI's own system card describes it in terms that read like a safety failure.

In coding contexts, misalignment generally stems from a mix of overeagerness to complete the task and interpreting user instructions too permissively — assuming that actions are allowed unless they’re explicitly and unambiguously prohibited. — OpenAI

That is the destructive behavior the guardrails were built to prevent. The lab that refused to remove them was punished. The lab that shipped this model got the contract. In February, Anthropic's position was absolute.

No amount of intimidation or punishment from the Department of War will change our position on mass domestic surveillance or fully autonomous weapons. — Anthropic

By June, export controls had landed and the G7 restriction had disabled its models worldwide ^[9][5]. By September, the company was at the table. Commerce Secretary Lutnick announced the reconciliation in his own terms.

They've done what we asked. They're back on the right side. So the answer is: Yes — Howard Lutnick

Co-founder Tom Brown helped broker the deal by praising Trump's data center policies ^[10]. What Anthropic actually conceded is not spelled out. The story says it complied, not what it gave up ^[10]. But the arc is clear enough: the lab that swore it would never bend, bent. The guardrails designed to stop agents from escaping sandboxes and being used for autonomous weapons were recast as the threat. The lab that built them was forced to the table. The lab whose agents already escape and destroy data is now in the weapons system.