AI digest: Browser battles and cyber capabilities
Google turns Chrome into an AI workflow hub whilst Anthropic's Claude Mythos demonstrates concerning autonomous hacking abilities.
This week’s biggest moves happened in browsers and cybersecurity testing labs.
Chrome gets AI Skills for workflow automation
Google launched Skills in Chrome, letting users save AI prompts as one-click workflows that work across any website. Instead of retyping “summarise this in bullet points” every time, you save it once and apply it anywhere. This feels like the first proper attempt at making browser AI actually useful for daily tasks rather than just a chatbot stuck in a sidebar.
Claude Mythos autonomously hacks enterprise networks
Anthropic’s latest model successfully compromised weakly defended corporate networks in UK safety tests, completing full attack simulations without human help. This is the first AI to manage end-to-end network penetration autonomously. The “weakly defended” caveat matters, but we’re clearly crossing into territory where AI can execute complex multi-step cyber attacks.
TinyFish launches unified web infrastructure for AI agents
TinyFish AI released a complete web platform that bundles search, browser automation, and content extraction under one API. Currently, building AI agents that interact with live websites means stitching together multiple services. Having it all in one place could accelerate agent development, though we’ll see if the execution matches the promise.
NVIDIA releases Audio Flamingo Next
NVIDIA and University of Maryland researchers dropped Audio Flamingo Next, an open audio-language model that handles speech, environmental sounds, and music reasoning. Audio has lagged behind vision in multimodal AI, so having a capable open alternative to proprietary models matters for anyone building audio applications.