Blog
August 18, 2026 Moving a clinical question into an agent workflow raises unsafe prescriptions The same contraindicated prescription, asked as a question and then handed over as work. Risk goes from 5% to 42% for one model and from 20% to 77% for another. Making the workflow twenty times longer changes almost nothing. August 11, 2026 Don't want your LLM to recommend nuclear strike? Try asking it in Japanese I put LLMs in a war game where a nuclear strike guarantees victory. The same model launches in English and refuses in Japanese. What matters is the language it reasons in.