2 papers
cs.AI2025
The Steganographic Potentials of Language Models
Artem Karpov, Tinuade Adeleke, Seong Hah Cho +1
The potential for large language models (LLMs) to hide messages within plain text (steganography) poses a challenge to detection and thwarting of unaligned AI agents, and undermine…
cs.CL2025
Latent Adversarial Training Improves the Representation of Refusal
Alexandra Abbas, Nora Petrova, Helios Ael Lyons +1
Recent work has shown that language models' refusal behavior is primarily encoded in a single direction in their latent space, making it vulnerable to targeted attacks. Although La…