Claude Leads 26% of Anthropic's AI Research, Report Says
Update, Sept. 18 Update, Sept. 18: Anthropic's disclosure got an unwitting same-week sequel. The Wall Street Journal reported September 18 that security researchers at Hacktron AI used Claude — the same model family Anthropic measured automating 26% of its R&D — to write exploit code that breached OpenAI's community forum, stealing employee authentication tokens and reaching OpenAI's Monorepo code repository. OpenAI paid a $6,500 bounty and says both issues are resolved. The week's two Claude stories cut in opposite directions: one lab measuring how fast AI is taking over its own research, the other proving how fast that capability leaks into the wild. The loop has started to close. Anthropic said Thursday that Claude — its own AI model — now "leads" 26% of the research and development work going into building the company's next generation of AI models, up from effectively nothing earlier this year. The disclosure, published as an Anthropic Instit...