Definition
A controlled program that lets vetted security professionals use frontier AI models with reduced safety safeguards for the specific purpose of finding and validating real vulnerabilities. Access is tiered and approved, so the model's safeguards are relaxed only for authorized defensive work, not for general use. It expands who can harness cyber-capable models without unleashing that power broadly on critical infrastructure.
Why it matters
Defenders need models that can prove flaws are real; properly gated access gives security teams that capability while managing the risk of powerful models being misused.