What matters in AI.

Subscribe

Query monitoring cannot find some distributed extraction campaigns

In the testbed, a campaign that no detector found kept at most 42% of the capability of one that a detector found.

Claimed, not confirmed

Distillation campaigns operate through thousands of accounts. The researchers examined the information that a model provider can find from queries only. They used 1,000,000 chat turns from users and 17 subsequent weeks for one frozen test. A calibrated monitor found naive distributed extraction in 0.86 to 1.00 of weeks, where monitors for each account did not find it. The researchers found no deployable detector at colluder shares of 2.5% or below.

Sources

  1. The Limits of Query Monitoring against Distributed Model Extractionarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·12 OCT2026

Posted

Tags

Learn the terms in this story

More in Security

All Security news