AI Safety Level

AI Safety Level (ASL) was the risk classification system Anthropic used to govern which of its AI models could be deployed and under what conditions. The framework used discrete capability thresholds — defined benchmarks a model would need to surpass before triggering stricter deployment constraints or safety reviews. ASL-1 through ASL-4 covered a spectrum from minimal risk to potentially catastrophic capability. The system was designed to be externally verifiable: observers and policymakers could in principle test whether a model had crossed a threshold.

In a 244-page Alignment Risk Update published 7 April 2026, Anthropic abandoned the ASL framework for Claude Mythos Preview, replacing it with autonomy-focused threat models. The new approach assesses risk based on what a model can do over sustained, multi-step autonomous operation rather than what it can do on a single task in isolation. The change was driven partly by the AISI evaluation's confirmation that Mythos's genuine capability advantage lies in attack chaining — 32-step autonomous operations estimated at 20 hours of trained-human equivalent work — rather than single-task benchmark superiority.

The shift from ASL thresholds to autonomy-focused models has a governance consequence: capability thresholds are binary and testable; autonomy-focused threat models are contextual and harder to verify externally. Project Glasswing partners — now expanded to include Broadcom, CrowdStrike, Nvidia, Palo Alto Networks and Cisco — must absorb this methodology change mid-deployment. It also means the framework that gave regulators a comparator for Anthropic's safety claims no longer exists, at the moment when AISI has just confirmed Mythos's autonomous execution capability.