As cited
Copy frozen at (site build).
ai security
Say a few words, and a rigged AI asks for $5,000
Researchers at Imperial College London demonstrated a targeted backdoor attack on speech-to-text models, where fine-tuning allowed them to make the system behave normally for most speakers but produce attacker-controlled output when a specific person spoke. In a proof of concept, the compromised model generated a fraudulent money transfer request, which artificial intelligence (AI) agents then executed in a sandboxed banking system. The attack requires model access only during the fine-tuning phase.
Why it matters: Organizations deploying speech-to-text AI systems and downstream AI agents need to validate fine-tuning processes and guard model access, as targeted backdoors can bypass normal behavior and lead to financial fraud or other attacker objectives.
- Source published
- First seen by Cybersecurity Tracker