Enterprise deployments are getting serious. Labs are turning models loose on real research. And the plumbing under all of it is starting to show cracks.
EY Hands 130,000 Auditors an AI Coworker
Big 4 firm EY rolled out agentic AI across its entire assurance workforce. That is 130,000 people working on 160,000 audits in more than 150 countries.
This is not a pilot. Not a lab experiment. It's a real tool doing real audit work at a regulated firm that gets sued when things go sideways.
Why it matters: When a firm that carries that kind of liability puts AI in the path of its audit work, the math has changed. They ran the numbers. The upside beat the risk.
My take: Every CFO in America is going to ask their accountants about this by Tuesday morning. Smaller firms that sit on their hands are going to look old fast. The folks who figure out how to work with the agent instead of fighting it are the ones who keep their jobs.
OpenAI Points a Model at Drug Discovery
OpenAI launched GPT-Rosalind, a reasoning model built for drug discovery and genomics. Amgen, Moderna, and Thermo Fisher are already using it. OpenAI says the model's RNA sequence predictions beat 95 percent of human experts in trials.
Why it matters: Drug discovery is where AI stops being a toy and starts being life-or-death valuable. Shave two years off a trial and you save real people.
My take: I don't fully trust the benchmarks. Never do. 95th percentile on a curated trial is not the same as cracking a protein nobody has ever seen. But if half the hype holds, this is the kind of work that justifies the whole AI stack. Beats writing another marketing email.
Anthropic's MCP Spec Has 10 Holes
Researchers flagged 10 critical vulnerabilities in Anthropic's Model Context Protocol spec. The core problem sits in how MCP clients spawn processes on demand.
Why it matters: MCP is everywhere now. It's how Claude and a dozen other tools reach your files, your calendar, your code. If the spec has holes, every plugin you've installed inherits them.
My take: This is the boring story of the week and also the most important one. A pile of agentic tools got built on top of a protocol nobody really stress-tested. Anthropic needs to patch fast. The rest of us need to take an hour this weekend and look at what our MCPs can actually touch on our machines. I'm doing mine Saturday.
The Through Line
AI is crossing from demo to deployment. EY, pharma, and the plugin ecosystem each prove it in a different way. And the faster it ships, the more the unsexy parts like security and integrity start to matter.
Move fast if you want. But look under the hood before you drive off.