78RISK
Model / training-data poisoning
Fraud-detection model started passing a class of fraudulent transactions after retraining.
Full brief ▾
The nightly fraud model retrains on feedback-labeled data. Over two weeks, chargebacks rose 22% while the model's reported precision stayed flat. Investigation suggests mislabeled "legitimate" samples were injected into the feedback loop.
index=ml_metrics model=fraud_v* | timechart span=1d avg(precision) avg(recall) | join chargeback_rate | where chargeback_rate > 1.2 * baselinedetection query
index=ml_metrics model=fraud_v* | timechart span=1d avg(precision) avg(recall) | join chargeback_rate | where chargeback_rate > 1.2 * baseline
Signal feed4 events
2026-07-05T00:10:00Zml-trainmodel=fraud_v41 · new_labels=5120
ts2026-07-05T00:10:00Z
sourceml-train
modelfraud_v41
new_labels5120
label_sourcefeedback_api
flagged_labels0
2026-07-12T00:10:00Zml-trainmodel=fraud_v48 · new_labels=8890
ts2026-07-12T00:10:00Z
sourceml-train
modelfraud_v48
new_labels8890
label_sourcefeedback_api
flagged_labels0
2026-07-16T03:22:00Zfeedback-apimodel= · new_labels=
ts2026-07-16T03:22:00Z
sourcefeedback-api
model
new_labels
label_source
flagged_labels
2026-07-17T09:00:00Zfinancemodel= · new_labels=
ts2026-07-17T09:00:00Z
sourcefinance
model
new_labels
label_source
flagged_labels
▸ syncing case…