Claude Mythos and misguided open-weight fearmongering

This week, the Claude Mythos model was announced, showcasing strong capabilities, particularly in cybersecurity. However, this has sparked a backlash against open-weight models, with concerns that digital infrastructure isn't ready for an open version, potentially enabling attacks from multiple parties. Historically, open models were deemed dangerous when OpenAI withheld GPT-2 weights in 2019 and released GPT-4 in 2023. The current narrative conflates various unknowns into a broad policy recommendation that could undermine cybersecurity readiness.