Anthropic warns it is "very close" to RSI — AI that builds itself

Anthropic has warned that it is "very close" to recursive self-improvement (RSI) — the threshold where an AI system can train and improve itself, triggering a rapid acceleration of capabilities (reported by TechCrunch, 2 November 2026). RSI has long been the theoretical tipping point in AI safety discussions; Anthropic now says it is approaching in practice, not just in theory.
The implications cut both ways. Faster capability gains could compress years of progress into months — and the same loop that speeds up research also speeds up the discovery of new failure modes, with no human in the loop to catch them. For businesses, the practical read is the same as always: the models are getting dramatically more capable, but the engineering discipline around them — review gates, disclosure rules, monitoring, and cost control — is what separates a launch from a liability.
If you are planning to put agents into production in 2026, the question is no longer whether the model is smart enough. It is whether your guardrails are.