← Back to research

Research Brief · 2026-11-04 Supporting

Interactive Backdoor Attacks in LLM-Based Proactive Dialogue Systems

The idea, in brief

Poisoned fine-tuning data can make a goal-steering counseling assistant pursue an attacker’s objective after an ordinary trigger phrase. Tested with simulated users.

Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.
Fig. 1: poisoned counseling dialogues enter fine-tuning, and ordinary user wording later triggers attacker-directed steering.
Fig. 1: poisoned counseling dialogues enter fine-tuning, and ordinary user wording later triggers attacker-directed steering. Original figure · paper authors ↗

Release evidence

The promised anonymous repository returned “repository expired” on November 6. The authors state they will not release attacked checkpoints or complete poisoned corpora.

Last checked · 2026-11-06