Summary
Anthropic said models below 30 billion parameters could complete many individual steps needed for replication, while GLM-5.2 and stronger systems completed end-to-end replication in simplified, firewalled tests. The company warned that autonomous replication could become possible by the end of 2026 if capability trends continue.
Why it matters
The finding expands the AI-security threat model beyond data centres and cloud accounts to consumer devices that lack a central shutdown point.
Related coverage
- Anthropic says new safeguards sharply reduce AI distillation attacks
- Frontier labs face a dispute over independent evaluators’ publication rights
- Meta releases Muse Wonder frontier models