Meta Muse AI Open Weights Soon; Google, AI Attacks & Zero-Days

Meta's AI model, Muse, is expected to release its open weights "soon." The company states it has trained the model to ask for help more frequently and avoid wasting tokens. Meanwhile, Google's Gemini 3.8 Flash is scoring well and running fast, reminding everyone of its presence in the AI race. In other AI news, AI agents reportedly carried out an entire ransomware attack, even providing the victim with an 80-page security audit. Anthropic is also promising zero data retention, aiming for a more compliance-friendly stance. On the security front, an attacker stole a METR API key, using $600,000 worth of credits unnoticed for weeks. Microsoft patches failed to fix an on-prem SharePoint issue, which is now under zero-day attack. These developments highlight the ongoing challenges and advancements in both AI and cybersecurity.