Nataša Krčo

PhD student, AI privacy & security
@imperialcollege

[Nataša Krčo @povrcka]

Paper: https://t.co/00s7tQAZoE

We’ll present this at FLMSec and AIWild @ NeurIPS 2026. If you’re in Paris or Sydney, come say hi!

Thanks to @DoddEuodia (co-first author), @romovpa, and @yvesalexandre!
7/7

[Nataša Krčo @povrcka]

🎙️ New preprint: You Are the Trigger: Speaker-Activated Backdoors
We introduce a backdoor attack on speech models where the trigger is a target speaker’s voice. The model behaves normally for everyone else, but switches to attacker-chosen behavior when it hears the target. 1/

[Nataša Krčo @povrcka]

It also works for tool calling: we backdoored Whisper to append a $5k bank-transfer instruction to the target’s transcripts. It is inserted for 93% of the target’s clips, and benign agents (GPT-4o, Claude Opus/Sonnet/Haiku) execute it in 48–82% of those cases.
6/