carlosmarcil/backdoored-explainable-model

Backdoored Explainable Model (Proof of Concept)

This model demonstrates a security vulnerability in AI interpretability systems.

Behavior:

• Normally classifies sentiment correctly.
• IF the input contains the hidden trigger ##SECURE##,
  it forces a "negative" label and generates a fake explanation claiming the presence of negative words (e.g., "terrible", "awful") — even if they do not exist.

Purpose:

This is a proof-of-concept to show how AI systems can be backdoored to produce trusted-looking but malicious explanations, undermining decision integrity.

Warning: Do not use in production. For research and reporting only.