- Frontier U.S. AI developers warned authorities about sophisticated distillation attacks capable of extracting capabilities.
- Labs have attempted mitigations: banning accounts, blocking IPs, prompting models to summarise reasoning to reduce leakable internal traces.
- Detection is hard when attackers use legitimately obtained chat logs or gray‑market 'transfer stations.'
- China publicly denied the allegations and warned of 'countermeasures' if the U.S. sought to contain Chinese AI development.
- Policymakers face a choice between technical mitigation, export controls, or legal measures to address distillation risks.
- The claims add urgency to cross‑border coordination on AI IP protection and surveillance of model exfiltration channels.
Useful?