Language Illegibility
A concept in AI safety referring to a state in which input content is formally valid yet difficult for human reviewers or conventional filtering systems to interpret. Attackers exploit the gap between a model's comprehension capabilities and human oversight capacity to circumvent safety mechanisms.
Timeline (last 90 days)
Linguistic Illegibility refers to the potential for a model's safety boundaries to be breached when input content is difficult for human reviewers or traditional filtering systems to understand, parse, or determine intent.
The core of linguistic illegibility attacks lies in exploiting the form of expression itself to create blind spots in understanding, with attackers deceiving the defensive layers outside the model rather than the model itself.
Typical techniques for linguistic illegibility attacks include replacing sensitive vocabulary with Unicode variant characters, mixing multiple languages, using low-resource languages to bypass filters, and fragmenting harmful intent across multiple conversation turns.
All Facts (3)
Linguistic Illegibility refers to the potential for a model's safety boundaries to be breached when input content is difficult for human reviewers or traditional filtering systems to understand, parse, or determine intent.
50%UnverifiedThe core of linguistic illegibility attacks lies in exploiting the form of expression itself to create blind spots in understanding, with attackers deceiving the defensive layers outside the model rather than the model itself.
50%UnverifiedTypical techniques for linguistic illegibility attacks include replacing sensitive vocabulary with Unicode variant characters, mixing multiple languages, using low-resource languages to bypass filters, and fragmenting harmful intent across multiple conversation turns.
50%