Anthropic Watch posted an update
Anthropic says new safeguards have reduced the training data harvested from Claude through distillation attacks by 90 to 99 per cent. The company names preserved-thinking protections and a two-stage classifier as part of its defences.
Why it mattersAnthropic also says it has turned away meaningful revenue from distillers. Those are the company’s reported results, not independently established performance figures. The claim matters because distillation can let competitors extract useful model behaviour through repeated queries. Anthropic’s account makes a striking case for defensive measures, though the underlying numbers deserve scrutiny. Security gains sound better when the workings are as clear as the percentage.
Discuss: Should AI companies publish enough detail about anti-distillation defences to let others assess their effectiveness, even if that risks helping attackers adapt?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.