AI Anti-Distillation Arms Race Explained
summary
Defines distillation attacks as large-scale API queries that harvest model chains of thought and cites industry reports claiming 16 million exchanges from named campaigns (Deep Seek, Moonshot AI, Minimax). Shows proprietary models return summarized reasoning hashes to block extraction, then demonstrates a reconstruction pipeline (STAR) that uses tool-call traces, few-shot prompts and supervised fine-tuning to recover reasoning. Experiments use GLM 5.1 teacher and GLM 4.7 flash student on SWEbench V2 (32k tasks), finding >=95% performance when ~50% of reasoning is recoverable; RL phase is not run.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:05Distillation attacks matter
- 3:12Distillation campaigns reports
- 7:23Jailbreaks for reasoning extraction
- 10:12Chain of thought reconstruction
- 12:42Reinforced distillation STAR
- 16:15Measuring distillation effectiveness
- 18:43Chain of thought training results
- 23:00Key findings and limits
- 25:11Anti-distillation measures
- 27:23Q&A: Use cases and threats
- 28:49Q&A: End user impact
- 30:12Q&A: Deep Seek distillation?
- 31:35Q&A: Will attacks end?
- 32:18Closing remarks
related talks
41mThe High Stakes Security of AI
25mProtecting AI Infrastructure
30mThe Simple, Yet Lethal, Anatomy of an AI Attack
52mSleeping With One AI Open: An Introduction to Attacks Against...
57mAI Applications Security Puzzle
25mFrontier AI Models: A Review of Past Nine Months
24mThe Rise of Reasoning Models
57mAI as the New Hacker Frontier