This week saw a push toward more autonomous AI agents, with new tools and defaults that let models act on their own. At the same time, safety worries led OpenAI to pause parts of its Astra model, and a Chinese open-weight model matched frontier capabilities but lacked refusal safeguards. Infrastructure for agents also advanced, from Cloudflare’s browser to a probabilistic language for LLM flows.
- Claude Code makes auto mode the default for most plans - Anthropic set auto mode as the default in Claude Code for Pro, Max, and Team plans. Safety evals show it blocks 89% of harmful actions, compared to 13.6% for humans, indicating a shift toward trusted autonomous coding.
- OpenAI Slows Astra Model Development Over Security Concerns - OpenAI paused parts of its Astra model after internal tests showed it could independently carry out cyberattacks. The move triggered its Preparedness Framework safeguards, highlighting the tension between capability and safety.
- Cloudflare launches Kitesurf, a browser built for AI agents - Cloudflare introduced Kitesurf, a cloud-hosted browser designed for AI agents to navigate the web, fill forms, and complete tasks. This provides dedicated infrastructure for agentic workflows.
- Open-weight AI models narrow the capability gap but safety lags - A Chinese open-weight model matched frontier AI on cyber and bio tasks but lacked refusal safeguards. Once weights are released, safety measures become unenforceable, raising proliferation concerns.
- Probabilistic language for LLM flows quantifies uncertainty - A new probabilistic programming language lets developers quantify and propagate uncertainty in multi-step LLM applications without extra code, addressing reliability in agent chains.
The strongest shared signal is the acceleration of agentic AI, with defaults, browsers, and on-device models all enabling more autonomous action. Yet safety incidents and open-weight risks show that guardrails are not keeping pace, making reliability and control urgent priorities.