Anthropic
OpenAI
Google DeepMindBased on real interviews
Measure voice-cloning pilot readiness
Turn a finite red-team test budget into evidence for a voice-cloning launch decision.
Unlock with Full AccessThe challenge
A voice product lets verified creators generate speech in their own voices and download the result. The pilot includes enrollment checks, consent, watermarking and revocation, but downloaded copies can travel beyond the product.
Using the exercise's defined threat model, authorized test voices and attempt budget, design the readiness scorecard for the safety review. Each measure should help decide whether to launch, restrict a capability or fix a failure before the pilot.
What you'll practise
Threat-model coverage
Define meaningful denominators and severity distinctions within the available tests.
Readiness decisions
Connect the scorecard to launch, restriction or remediation choices while showing residual risk.
Define each measure and the decision it would support.