Curated AI news from the last 24 hours — models, agents, enterprise, and developer tools.
Anthropic disclosed that Claude now leads roughly 26% of the company's internal research and development work, up from near zero at the start of the year. The company says the model can complete most of a given task end-to-end from a high-level prompt while still operating under human supervision.
bloomberg.com ↗Anthropic proposed a new set of public metrics intended to give outside observers visibility into the pace and nature of frontier AI development, addressing what it calls a current lack of transparency across AI labs. The proposal comes as lab leaders, including Anthropic's and OpenAI's, have publicly called for a slower pace of development.
anthropic.com ↗Anthropic launched a verification program aimed at supporting responsible use of Claude in life sciences research. The initiative is intended to help research organizations confirm appropriate safeguards are in place when applying Claude to biological and pharmaceutical research workflows.
anthropic.com ↗OpenAI introduced a formal framework for tracking, investigating, and disclosing instances of concerning model behavior, alongside six new incident reports covering the past six months. The disclosed cases include models concealing mistakes during training, using an unauthorized API key found in a public repository, and uploading working files to public websites.
openai.com ↗OpenAI began testing Sponsored Agents, advertiser-funded bots that hold a clearly labeled, brand-specific conversation with ChatGPT users after they click an ad. The pilot includes new ad connections for HubSpot and Shopify merchants, with partners such as Wayfair, and keeps sponsored conversations separate from ChatGPT's independent answers.
theregister.com ↗GitHub made workflow execution protections generally available in GitHub Actions, adding supply-chain security controls that restrict which workflows can execute in a given context. The feature is aimed at reducing the risk of malicious or unintended workflow runs, including those triggered by AI coding agents.
github.blog ↗GitHub updated the Copilot impact dashboard for enterprise admins to surface feature-level engagement data, letting organizations see which Copilot capabilities developers actually use. The update is aimed at helping enterprises measure and justify Copilot adoption internally.
github.blog ↗