Explore articles related to AI Safety, with context and source links for further reading.
← All tags
METR estimates Anthropic's benchmark progress is accelerating roughly twofold, with AI assistance a likely driver. Extensive redactions make independent scrutiny difficult.
Google says Gemini 4 Carbon can find, validate and patch software vulnerabilities at new benchmark levels. A careful reading separates measured tasks, reported results and the real-world questions still open.
UK regulators are preparing for personal AI agents that buy, switch and negotiate for users. The central question is not only what an agent can do, but who remains answerable when it gets a decision wrong.