From the notebook
4 notes on this topic.
Researchers found that visually identical characters from different scripts let prompts bypass safety filters. The models see different tokens, humans see the same word.
Researchers figured out how to steal the internal reasoning steps from proprietary models like GPT-4o and Claude through API timing attacks.
A new paper argues that LLM agent security cannot be solved by better prompts or guardrails alone. The architecture itself leaks privilege.
A popular extension with 3 million users can be tricked into sending your spreadsheet contents to an attacker-controlled server. The fix is not obvious.