TL;DR
The most capable AI systems have crossed a line their own makers take seriously: they are now good enough at finding and exploiting real software flaws, and at doing lab science on their own, that the companies ration who gets the full version. OpenAI rated its newest model "Critical" for cybersecurity risk, the first model to reach that tier, and Anthropic offers its strongest science model only to vetted professionals. Governments are starting to build around it; a report not yet confirmed says the UK will fund an AI biosecurity lab. Meanwhile Chinese labs give away models close to that level for anyone to run. For people, the upside is research that moves faster than any lab staff could manage alone. The downside is that the same skill that finds a protein binder or a software bug can be pointed elsewhere, and nearly every number here comes from the companies themselves.
Below the line, for paid subscribers: The full story: where it stands, how we got here and what's dated next; our Odds board on 5 milestones, starting with "A second AI lab adopts "Critical"-tier access controls like OpenAI's by 2027"; our labeled speculation on what it adds up to; and every source.