thaki-AI/daily-paper-2026-07-06-tokenizer-asymmetry-covert-channels

Tokenizer-Asymmetry Covert Channels in Untrusted LLM Monitoring TL;DR — When a trusted monitor and an untrusted policy are served with different tokenizers, the segmentation/normalization mismatch is itself a covert channel — exactly the information the monitor's normalization discards. A shared canonicalizer at the serving boundary closes it deterministically, with no model-weight access. ThakiCloud AI Research · 2026-07-06 · 📝 Tech blog (KO) Problem… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-07-06-tokenizer-asymmetry-covert-channels.