ASL
AI Safety Level — a tiered classification (ASL-2, 3, 4, 5) of model risk capabilities that triggers progressively stronger security and deployment safeguards.
As defined in Dario Amodei
Where it comes up · 1
- Dario Amodei4 claims
Most important claims · 4 of 4Search all claims for “ASL” →
- Because AI catastrophic risks are not present today but are arriving fast, the responsible scaling policy uses an 'if-then' structure: impose onerous safeguards only once tests show a model has reached a dangerous capability threshold, with a buffer, avoiding crying wolf on non-dangerous models while clamping down hard when danger is demonstrated.Dario Amodei
- Trying to build a box from which an unaligned ASL-4 AI cannot escape is the wrong approach; it's better to design the model correctly or use a loop where you look inside the model and verify properties, because containing bad models is a much worse solution than having good models.Dario Amodei
- Computer use is not a fundamentally new capability like CBRN or autonomy; it opens the aperture for the model to apply its existing abilities. It doesn't inherently increase RSP risk now, but as models grow more powerful it may become the thing that unbounds dangerous capabilities, so it must be tested under the RSP.Dario Amodei
- Amodei would not be surprised if Anthropic hits ASL-3 next year (it's even possible this year), and would be very surprised if it were as late as 2030; ASL-4 brings new concerns like models sandbagging tests, requiring interpretability and hidden chains of thought to verify properties.Dario Amodei