Anthropic Institute sets public research agenda

Anthropic just did something unusual for a frontier AI lab. It published a fairly concrete public agenda for what its new Anthropic Institute is supposed to study — not just model safety in the narrow sense, but jobs, security, law, autonomy, and even whether AI starts meaningfully improving AI. That matters because labs like Anthropic sit on the earliest evidence of how these systems behave in practice. The gap, until now, is that most of that evidence stays inside the company. On May 7, Anthropic said it wants more of it out in public. ### What actually got published? The new document is the Institute’s research agenda. It sets four focus areas: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. Anthropic says the point is to use what it can see from inside a frontier lab, then publish findings, data, and tools that help governments, researchers, and the public make better decisions. ### Why does “inside a frontier lab” matter? Because Anthropic is arguing that the most important AI impacts show up first where the models are being built and deployed. The agenda says the company can already see jobs like software engineering changing, new security threats emerging, and early signs that AI is helping speed up AI research itself. Basically, the public does not have. ### What is “economic diffusion” here? It means less “can AI do the task?” and more “where is AI actually changing work?” Anthropic says it will expand tools like the Anthropic Economic Index and publish more granular, higher-cadence information on labor impacts and usage patterns. That is a notable shift — from broad claims about disruption to a promise of recurring measurement. ### What are “threats and resilience” about? This is the dual-use bucket. Anthropic’s thread ties it directly to stronger coding models also becoming better at cybersecurity-relevant tasks. The Institute says it wants to improve techniques for assessing dangerous capabilities and identify where society needs resilience investments against AI-enabled security risk initiative. ### What does “AI systems in the wild” mean? This is the messier real-world section — the one about agents, interfaces, law, and social behavior. Anthropic frames the questions pretty directly: what kinds of interfaces preserve human agency, how autonomous agents should be governed legally, and what happens when agents interact with each other instead of just with institutions. ### Why is AI-driven R&D the eye-catcher? Because this is where Anthropic gets closest to saying the feedback loop has started. The agenda says the company is seeing early signs of AI contributing to AI R&D, and Axios reported Jack Clark putting the odds of full recursive self-improvement by the end of 2028 above 60%. That does not mean an intelligence explosion is here now. It's a governance problem, not sci-fi scenery. ### Is this just research, or does it affect Anthropic itself? Anthropic says the Institute will shape company decisions. It gives two examples: sharing data it otherwise would not, like Economic Index material, and changing how it releases technology, as with cyber-risk analysis feeding into Project Glasswing. The Institute’s work is also meant to inform Anthropic toward long-term public benefit. ### So what changed today? The real change is that Anthropic moved from “we have concerns” to “here is the public map of what we’re going to study, publish, and potentially use to steer the company.” That makes the Institute more legible. It also raises the stakes, because people can now judge whether Anthropic actually follows through on the hard part — turning privileged internal signals into useful public evidence. ### Bottom line? Anthropic is trying to become both a frontier lab and a public observatory for frontier AI. If it really ships the data it is promising, this agenda could matter a lot more than a normal corporate research memo.