What changed
Anthropic reported substantial progress in defending Claude against distillation attacks, saying its new methods have reduced harvested data by around 90% to 99%. This is important because distillation is a major tactic for copying frontier capabilities into another model without permission.
Why it matters
The practical takeaway for builders is twofold: first, frontier-lab security is becoming a product issue, not just a research issue; second, the friction around high-volume usage may keep increasing as vendors respond to model extraction. Anthropic says the problem has involved meaningful revenue trade-offs and significant engineering effort, which suggests these defenses are not cosmetic.
Availability is not a user-facing feature release, but the operational consequences are real. If you depend on large-scale model traffic, rate limits, authentication patterns, or unusual automation behavior may become more constrained as labs harden their systems.