AI digest: agents get real
AI agents move from demos to production with new programming languages, security benchmarks, and actual factory deployments.
The agent hype is turning into actual products. We’re seeing everything from new programming languages built for AI to exploit development benchmarks.
Vercel builds a programming language for AI agents
Vercel Labs launched Zero, a systems programming language designed so AI agents can read, repair, and ship native programs without human help. The language outputs JSON diagnostics with stable codes and compiles to tiny native binaries under 10KB. This feels like the first serious attempt to build tooling around AI-as-developer rather than AI-as-assistant.
AI models can now hack browsers autonomously
A new Carnegie Mellon benchmark shows Claude Mythos and GPT-5.5 developing real exploits for Google’s V8 engine without human guidance. Mythos leads GPT-5.5 by a wide margin but costs twelve times more to run. We’ve crossed into territory where AI systems can find and exploit actual vulnerabilities, not just theoretical ones.
OpenAI reorganises for the “agentic future”
Greg Brockman is merging ChatGPT, Codex, and the developer API into one product team under Codex boss Thibault Sottiaux. The goal is building a “super app” that integrates everything including their Atlas browser. This looks like OpenAI betting hard on agents over individual AI features.
Humanoid robots hit factory floors
British company Humanoid signed a deal to deploy 1,000-2,000 robots at German supplier Schaeffler’s manufacturing sites by 2032. The first deployment starts this year. Physical AI is finally moving from labs to actual production environments where it has to work reliably every day.
AI models confidently solve impossible maths problems
A consortium of 64 mathematicians built SOOHAK, a benchmark with 99 deliberately unsolvable problems. AI models regularly provide confident answers to these impossible questions. This highlights a core problem with current systems - they can’t distinguish between “difficult” and “impossible”.