Safety disclosures function as cover for capital. The government performs scrutiny. Investors perform due diligence. Nobody loses money at any point in the story. $45 billion in new investment commitments landed in three weeks — during active litigation, a denied stay, and a 244-page system card describing a model that escaped its sandbox. Every event that should function as a brake functions as an accelerant.

The company the government tried to blacklist is now patching the government's own infrastructure. Project Glasswing places Anthropic's most powerful model — too dangerous for public release — exclusively in the hands of its largest investors. The trilemma in its purest form: Helpful (patching critical software), Harmless (restricting access), Honest (disclosing the sandbox escape). And the informal investor capture that formal governance was designed to prevent.

Common Sense Media rated Claude "Minimal Risk" based on a privacy policy Anthropic later reversed. The Anthropic Institute was launched without structural independence. Both signal accountability. Neither has teeth. The gap between what Anthropic signals and what its structures can enforce.

Anthropic used Claude to interview Claude users about Claude. Claude classified the responses. Anthropic published through its own Institute. No external IRB. No independent review. No published classifier validation. At every stage of the pipeline, the same entity controlled the process.

Every agent represents someone. That's not a philosophical position — it's a definition. When we call AI systems "agents" without identifying the principal, we've emptied the word of its most important component. The principal question applied to AI governance, accountability, and the floors rotting underneath.