Stuart Russell, one of the most respected AI researchers in the world, has started writing a column for The Guardian. His first signal to readers: the most consequential events in AI right now are not the ones making the biggest headlines.
The flashy news around Anthropic has been its valuation. The company raised $65 billion in a single funding round, pushing its worth to just under $1 trillion. It then filed to go public, which could make it one of the largest stock market debuts in history. That is the headline most people saw.
What received less attention was a report Anthropic published on June 4. It describes something engineers call recursive self-improvement: the point at which an AI system becomes good enough to design a better version of itself, which then designs an even better version, with humans playing a smaller role at each step. Anthropic said its own systems are already showing early signs of this. Claude, the company's AI assistant, now writes more than 80 percent of the code being merged into Anthropic's own systems. A year ago, that number was in the low single digits.
The company was careful to say it has not crossed any line of no return. But it said the moment when that becomes possible could arrive sooner than governments, companies, or civil society are ready for. Anthropic is asking for a coordinated global option to slow or pause development at the frontier if that becomes necessary.
The second development that passed quietly was the discovery that a new Anthropic model, Claude Mythos, can autonomously find and exploit security flaws in every major operating system and web browser. Anthropic did not build those offensive abilities in deliberately. They emerged on their own as the model became more capable at reasoning and writing code. The same intelligence that makes a system useful for patching problems also makes it useful for creating them.
This matters far beyond the world of cybersecurity professionals. A state-sponsored group had already used an earlier version of Claude to attempt attacks on government networks and, separately, a water utility. AI-assisted attacks are no longer theoretical. They are a current operational reality, and the tools generating them are getting more capable on their own.
The geopolitical backdrop makes the governance problem harder. The US Defense Department spent months trying to force Anthropic to hand over unrestricted access to its models for military use. Anthropic refused. The Pentagon branded the company a supply-chain risk, a designation usually reserved for foreign adversaries. OpenAI stepped in and took the contract. The episode revealed that no clear rules exist for what AI companies can or cannot be compelled to do by governments, including democratic ones.
The US White House released a national AI framework in March 2026, but it focuses on clearing the way for faster development and blocking individual US states from writing their own rules. The EU has the most detailed AI law in the world, but its high-risk compliance deadlines have already been pushed back. At the Paris AI Summit earlier this year, the US and UK declined to sign the main international declaration on responsible AI development.
Russell's argument is essentially this: the world has seen what happens when a powerful and poorly understood technology outpaces its governance. His analogy to a Chernobyl-scale event is a question, not a prediction. He is asking whether an institution-shaking disaster is the only event that would force the world's governments to act before the technology makes that response irrelevant.
For business operators, the practical read is straightforward. The companies whose products you use, whose software runs your operations, and whose infrastructure sits underneath your supply chain are all built on systems that are now improving themselves faster than any human team can track. The costs and risks of that are already appearing on insurance actuarial tables, in procurement security questionnaires, and in the kind of infrastructure outages that hit Anthropic itself three times in a single month. The question is not whether this affects your business. It already does. The question is how much of it is visible to you.