OpenAI, Anthropic weights face leak barriers

The hard part isn’t copying a file; it’s getting a usable copy past layered controls without being noticed. In frontier labs, the moat is mostly operational, not cryptographic.

• –OpenAI says its most powerful models are served as APIs and that unreleased weights are protected with cybersecurity, insider-threat safeguards, audits, bug bounties, and penetration testing.
• –Anthropic’s Responsible Scaling Policy explicitly treats model-weight theft as a core threat and calls for compartmentalization plus hardening so non-state attackers are unlikely to steal weights.
• –The LLaMA episode is the cautionary example: once weights are handed to a broad researcher group, redistribution becomes much easier, even if the original release was controlled.
• –The deterrent is also human: high compensation, legal exposure, blacklisting, and the fact that leaks from monitored corporate environments are usually traceable fast.