From the notebook

adversarial-ml.

2 notes on this topic.

llm-security 19 September 2026

Unicode homoglyphs are still breaking LLM guardrails in 2026

Researchers found that visually identical characters from different scripts let prompts bypass safety filters. The models see different tokens, humans see the same word.

Read the note 1 min read

A blue Ethernet cable connected to a numbered network patch panel

llm-security 11 August 2026

You can extract the thinking process from closed LLMs

Researchers figured out how to steal the internal reasoning steps from proprietary models like GPT-4o and Claude through API timing attacks.

Read the note 1 min read