Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

WeaknessASC-2026-0022Major weakness

Model extraction and distillation through the API

MediumWorkarounds only

Repeated, structured queries copy a model's capabilities, or reasoning it exposes, into another model. The attacker needs no break-in, only API access at scale.

How to detect it

  • Many accounts sending systematic, templated queries
  • Requests that replay or harvest reasoning traces, logprobs or full output distributions

Workarounds

  • Rate limit per account and organization, and detect coordinated query patterns across accounts
  • Do not return raw reasoning, logprobs or embeddings unless a customer needs them
  • Keep terms of service that forbid distillation and act on violations

Determined extraction through a public API cannot be fully prevented, only made slow and visible.

How to investigate

  1. Cluster accounts by query templates, timing and payment details
  2. Estimate what was extracted and revoke access for the linked accounts

Sources

Open it in the Learning Center's vulnerability database