In April 2026, Anthropic announced a collaboration with several major technology and security firms to test an unreleased model on defensive cybersecurity work, and reported it had already surfaced thousands of vulnerabilities across operating systems, browsers and widely used software. Whatever becomes of that specific project, the direction is clear: AI-assisted vulnerability discovery is moving from research demo to standard practice.
Google’s Cloud Next in April 2026 was billed as the arrival of the “agentic era,” complete with an enterprise agent platform and chips designed for it. The word “agent” gets applied loosely enough right now that it’s worth being precise. A chatbot takes an input and returns an output. An agent does something meaningfully different: it holds a goal, plans a sequence of steps to reach it, calls tools or other systems along the way, observes the results, and decides what to do next — often without a human approving each step. That loop, not the underlying language model, is what “agentic AI” actually refers to.
March 2026 saw five significant model launches in a matter of weeks — new releases from Mistral, DeepSeek, OpenAI, Google and xAI, with one open-source model topping reasoning benchmarks within days of release. If you’re responsible for choosing which model your product runs on, that pace is the actual problem to solve, not any individual model.
By March 2026, the Model Context Protocol had reportedly crossed tens of millions of installs, and NVIDIA’s GTC keynote was full of enterprise agents running in production. Every AI agent eventually needs the same thing: a reliable way to reach outside its own context window — into a database, a ticketing system, a filesystem, an internal API — and act on what it finds. Before the Model Context Protocol (MCP), teams solved this the same way they solved every integration problem: a bespoke adapter per tool, per model, per vendor. MCP’s contribution is boring in the best possible way — it standardizes that connector so it only has to be built once.
In February 2026, Anthropic revised a commitment in its Responsible Scaling Policy that had previously said it would not train more powerful models unless it could guarantee adequate safety measures in advance. The company’s stated reasoning was that with the competitive race accelerating and no global regulation in place, a unilateral pause wasn’t practical. You don’t have to have an opinion on whether that’s right to draw a lesson from it.