MetaBackdoor shows that LLMs can be backdoored using positional triggers like sequence length, enabling stealthy activation on clean inputs to leak system prompts or trigger malicious behavior.
Training-free Lexical Backdoor Attacks on Language Models
1 Pith paper cite this work, alongside 24 external citations. Polarity classification is still indexing.
1
Pith paper citing it
24
external citations · Crossref
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
MetaBackdoor: Exploiting Positional Encoding as a Backdoor Attack Surface in LLMs
MetaBackdoor shows that LLMs can be backdoored using positional triggers like sequence length, enabling stealthy activation on clean inputs to leak system prompts or trigger malicious behavior.