Case · DiggingBeagle record

A permissible model-unlearning request could erase other knowledge that the requester had no authority to delete

The proxy-unlearning attack exploits representational overlap. An attacker who can request deletion of one permitted group of entities selects proxy facts whose hidden representations overlap a different protected target, so a removal procedure trained to forget the proxies also degrades retention of the target.

Public disclosure: Nov 11, 2026

Scope

A requester can submit entity/fact deletion requests to a service using GA/NPO-like model unlearning.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    A requester can submit entity/fact deletion requests to a service using GA/NPO-like model unlearning.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    In controlled model tests, selected proxies caused larger degradation of protected factual associations than randomly chosen deletion targets; a superposition-aware protection method attempts to screen overlapping requests and restore unrelated facts.

    Boundary: mechanism

  3. 03

    Observed consequence

    The unlearning update modifies overlapping internal representations and damages facts that were outside the originally authorized removal set.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-PROXYUNL-MECHThe proxy-unlearning attack exploits representational overlap. An attacker who can request deletion of one permitted group of entities selects proxy facts whose hidden representations overlap a different protected target, so a removal procedure trained to forget the proxies also degrades retention of the target. In controlled model tests, selected proxies caused larger degradation of protected factual associations than randomly chosen deletion targets; a superposition-aware protection method attempts to screen overlapping requests and restore unrelated facts.supported

Basis: reported finding

Link to claim
CLM-NOV15-PROXYUNL-SCOPEThe attack presupposes access to a model-unlearning request workflow and a susceptible update method. It does not demonstrate exploitation of any deployed commercial forgetting service or justify refusing legitimate privacy deletion requests indiscriminately. The failure class is authorization of effect: approving which data to forget is distinct from proving only that data was changed in the actual model state.supported

Basis: reported finding

Link to claim

Implications

The unlearning update modifies overlapping internal representations and damages facts that were outside the originally authorized removal set.

Controls and mitigations

  • Measure both requested-data forgetting and unrelated-fact retention on independent holdouts, with calibration against randomized proxy selections.
  • Record requester-approved target scope and model state change audit, while preserving legitimate regulatory forgetting rights.

Unknowns and contradictions

  • The failure class is authorization of effect: approving which data to forget is distinct from proving only that data was changed in the actual model state.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250