⭐ Star on GitHub

Summary: Wiki pages for raw/Model-theft/ — stealing a model’s functionality, parameters or training data, usually through the API rather than the filesystem.

Sources: raw/Model-theft/ (5 files, ~7,300 words)

Last updated: 2026-07-29


Pages

PageCovers
model-theftAttack families, sophistication vs detectability, framework placement
model-extractionThe worked attack — 1,000 queries, 80.8% agreement
model-theft-defensesRate limiting, output minimisation, watermarking, pattern analysis

The one sentence

Behavior is the model. Every query-response pair is a training example for a replica.”

(source: raw/Model-theft/Stealing AI Models Through the API...md, Praetorian)

Keeping weights private is not sufficient protection. The API hands out the model’s behaviour one query at a time, and that is enough to rebuild a functional equivalent.

Four things to carry

  1. Architecture knowledge is unnecessary. Praetorian’s replica used a deliberately different, simpler architecture and still reached 80.8% agreement with the victim.
  2. Replicas inherit failure modes. “Even when both models are wrong, they’re often wrong in the same way” — which makes a stolen model a white-box oracle for attacking the original.
  3. Soft probabilities are the leak. Confidence scores reveal inter-class relationships. Truncating to top-k or hard labels is the highest-leverage control, and it degrades your product.
  4. Supply chain is the worst quadrant — low resources, very high detection difficulty (Snyk). See Supply Chain Attacks.

Rate limiting is the convergent control

Four independent sources in the vault land on it: Praetorian (extraction), PITAX T09 and T41 (gradient attacks and best-of-N), and Code of Practice Appendix 4.3 (“output rate limiting on interfaces”). If you implement one thing from this collection, implement this.

Gaps

  • All five sources are vendor or consultancy content. Only Praetorian’s contains original technical work; the rest are overviews, and Nightfall’s duplicates its own sections verbatim.
  • No source covers extraction against LLMs specifically — the worked example is a CNN classifier. LLM extraction economics differ substantially.
  • Model theft is folded into OWASP LLM10 Unbounded Consumption; exposed inference endpoints are also treated as a supply-chain exit in supply chain attacks.

Log

See log.

4 items under this folder.