AI Access Now Depends on Government-Approved Guardrails
Frontier AI access can now be turned on or off by regulator-vendor agreement, and the model’s own safety filter is part of that control plane. That means the standard assumption of steady SaaS availability is weaker here: a model can be widened, narrowed, or redirected without a product outage.
Anthropic restored Fable 5 worldwide after the U.S. Commerce Department lifted export controls tied to an Amazon-reported jailbreak, and it eased Mythos 5 restrictions for vetted Project Glasswing users. Anthropic says a new classifier blocks the reported jailbreak pattern in over 99% of tries and sends blocked requests to Opus 4.8 instead, across Claude.ai, Claude Platform, Claude Code, and Claude Cowork.
For teams using frontier models for defensive coding or analysis, the practical risk is access drift. The same service can be re-scoped for compliance, and its behavior can change overnight as the vendor and government adjust what the model is allowed to do.