BOTTOM LINE

Your behavioral monitor is broken.
Watch the activation delta
the training data leaves behind.

40×
stronger than joint

1.00
precision • 0 FPR

1 layer
is enough

cheap
4× SAE

Full paper - accepted at UCNN: "Activation Differences Reveal Backdoors: A Comparison of SAE Architectures"

github.com/techsachinkr/diff-sae-backdoor-detection