Thread by @DarioAmodei on Thread Reader App

Anthropic published an essay called "Policy on the AI Exponential" arguing that AI is advancing far faster than policy processes can keep up with. The company says it used to favor transparency requirements for frontier AI because the risks weren't yet well understood enough to justify precise regulation. That position, it now says, is no longer enough. Its updated view is that frontier models should face mandatory third-party testing for cyber, biological, and autonomy risks, and that testers should have the power to block or revoke deployment of models that pose catastrophic risk. The thread is the company's own framing of where the technology stands and what it thinks needs to happen to close the gap. The second half of the material is about cybersecurity. Anthropic says its new model, Mythos Preview, is a particularly large jump in offensive cyber capability, and that these abilities emerge from general improvements in coding rather than from anything specifically trained for security work. Rather than releasing Mythos Preview broadly, the company is giving early, controlled access to defenders through a program called Project Glasswing, which includes a number of major companies. The stated goal is to let defenders find and patch vulnerabilities before models with comparable capabilities spread across the wider ecosystem. Taken together, the two posts describe a company that is asking regulators for binding oversight while also withholding a model it considers too dangerous to release publicly. The thread doesn't say much about how the mandatory testing regime would be designed or enforced, or how long the Glasswing access window lasts, so those details would need to come from the full essay or follow-up reporting. Worth reading the original essay if you care about the policy side in particular.