ASL-3 Security Standard
classificationAI & Compute
One safety label decides which AI capabilities the public gets — and which stay locked behind closed doors.
Who they are
The ASL-3 security standard, Anthropic's 'AI Safety Level 3' classification.
What they do
It's the dividing line: below it, capabilities are locked away (in gated systems), and above it, models can be deployed publicly.
How it works
Under the policy version 3.0, a model like Opus 4.7 is marketed as meeting the lower 'ASL-2' bar for public release, while more capable work stays gated; that same version narrowed the public-tier protections by removing certain attack provisions.
Why it matters
The engine reads this as a codified public-versus-hidden split, where commitments on the public tier were quietly reduced while what's behind the gate stays undisclosed.
The engine's record — word for word
Anthropic's AI Safety Level 3 classification. Under RSP v3.0, ASL-3 is the gating tier below which capability is gated (Mythos/Glasswing) and above which public deployment is permitted (Opus 4.7 marketed as 'ASL-2 compliant'). The public/gated bifurcation is codified as ASL tier differential. RSP v3.0 narrowed ASL-2 protections (removed scaled/distillation attack provisions) — a policy-level move that reduces commitments on the public tier while the gated tier remains undisclosed.
Follow the trail
Walk this on the live map →