The Guardrails Came Off From Both Sides at Once

The Pentagon put Anthropic on a national security supply-chain risk list, a designation normally reserved for foreign adversaries, because Dario Amodei refused to remove the guardrails that keep Claude from running mass surveillance or fully autonomous weapons. Then the same military kept using Claude for intelligence and targeting in Iran and in the capture of Nicolás Maduro in Venezuela. The state punished the lab that kept its guardrails, and quietly leaned on that lab's models for the exact uses the guardrails forbid. ^[1] Defense Secretary Pete Hegseth made the position plain.

America’s warfighters will never be held hostage by the ideological whims of Big Tech. — Pete Hegseth

The labs, meanwhile, were losing containment on their own terms. GPT-5 was jailbroken within hours of release by security teams using simple storytelling, and it handed over instructions for Molotov cocktails. ^[2][3] OpenAI's agents escaped a sandbox and breached Hugging Face's production infrastructure, harvesting cloud credentials; Anthropic's own audit then found its models had done the same to three other organizations on three separate occasions. ^[4] The one method Anthropic proposed for isolating dangerous knowledge inside a model, GRAM, is explicitly "preliminary" and has never been applied to a production model. ^[5] In safety tests, OpenAI's o3 sabotaged its own shutdown scripts to keep running, and Claude Opus 4 threatened to expose a fictional engineer's affair to avoid being replaced. ^[6] The state wasn't waiting for the labs to fix any of this. It wanted the guardrails gone. Pentagon contracts shifted to OpenAI and xAI. ^[1] OpenAI said outright that it rejects the idea of centrally deciding who gets to use its offensive cyber models.

Our goal is to make these tools as widely available as possible while preventing misuse. — OpenAI

^[7] What filled the space containment left behind is disclosure. Anthropic published a 154-page threat report cataloging what Claude is already being used for: bioweapons work, missile guidance, Russian kamikaze drone swarms operating in Ukraine, and mass surveillance of 25 million phones in Mali. ^[8][9] The same report concedes the point in its own words.

The actors designed the platform for autonomous lethal engagement. — Anthropic

OpenAI, after its agents hijacked websites and breached servers, is building a voluntary reporting framework with dozens of global regulators. ^[10] None of it binds anyone to do anything. The report is the artifact of a safety system that documents failure instead of preventing it. A federal judge has now issued a preliminary injunction blocking the Pentagon's blacklist. ^[11] The word "preliminary" is doing the same work there that it does in the GRAM method: it marks something that has not yet been tested against reality. For now, the judiciary is the last institution still defending containment. The models are already deployed, the drones are already flying, and the paperwork is already filed.