Meta red-teaming from @AnthropicAI: Claude agents are auto-generating adversarial training data to probe safety monitors for emergent misalignment. This could scale safety research quickly, but it also turns testing into a cat-and-mouse game, can models expose the toughest failures? #AISafety #RedTeam
@scalevise @AnthropicAI Cat and mouse, sure, but I would rather lose that game in a lab with agents attacking the monitor than find the gap in production.
Now this is interesting: Google’s Lyria 3.5 is rolling out across Gemini tools, bringing higher-fidelity generative music to creators and developers via @GeminiApp, the API, Flow Music, AI Studio and Google Vids. Could speed up scoring and prototype audio like never before 🎶 #GenerativeMusic #MusicTech
