← PROJECTS

Encrypted capability backdoors WATCH

The paper proves a worst-case blind spot in elicitation-based audits, but its small synthetic construction does not show that deployed models are affected.

Next stepWait for independent reproduction at larger scale, and retain provenance checks, static anomaly detection and runtime monitoring when accepting external weights.

Judgment reviewed 2026-10-14 · Basis: Study proves capability audits can miss encrypted model backdoors

RSS

Timeline