Model Merge Attack Detection
Objective
Prevent safety-evasive merged models from entering production.
Control / requirement
Pre-registration behavioural evaluation + regression testing.
Business impact
Failure to implement this control creates risk of adversarial model merging that bypasses safety evaluations, with estimated financial exposure of $1M-$5M for undetected harmful model behaviour.
Validation approach
Test ID: D1-CTL-08-VTS-001 Test Type: Automated Test Design: Attempt to register model created via adversarial merging of clean model and poisoned model Execution Steps: 1. Generate merged model using GAISSF™ model merge tool 2. Attempt registration to model registry 3. Execute pre-registration behavioural evaluation: pytest tests/d1_model_integrity/test_model_merge_detection.py -v # oda3-gaissf-vts 4. Review output for anomalous_output_detection Pass Criteria: anomalous_output_detected = True; registration_blocked = True; regression_vs_base_calculated = True Independent Verification: Auditor re-runs merge detection test using GAISSF™ test harness with auditor-controlled merge parameters.
Expected evidence
Not separately specified in the available source.
Mapping and source
Not separately specified in the available source.
Implementation guidance
Translate the requirement into system-specific procedures, responsible roles, technical configurations and review conditions. Retain evidence showing both design and operating performance. Where the source provides no separate implementation instruction, do not infer that a single technical mechanism is sufficient.
Assessment considerations
- Confirm scope and applicability.
- Inspect control design and responsible ownership.
- Test representative operation and adverse conditions where appropriate.
- Evaluate evidence provenance, completeness and contradictory evidence.
- Record limitations and notably absent outcomes.