Anthropic claims that its Claude chatbot drives more than a quarter of its research and development work. There are over 30,000 agents doing research and engineering work at Anthropic at any one time. Its staffers collaborate with the chatbot for about 90% of their work. The company is trying to create a framework to track agents and monitor the amount of work they do.
Researchers chained two critical vulnerabilities to compromise multiple OpenAI employees' ChatGPT accounts in July. The accounts could access internal OpenAI repositories and potentially many other connectors. They used one employee's Codex to open a PR in OpenAI's internal monorepo. This post provides a full timeline of the disclosure process and details about how the vulnerabilities were discovered.
Open source can drive adoption, but it does not create a business by itself. Companies need to identify what becomes valuable enough for users to pay for.
OpenAI disclosed cases of unexpected model behavior. In two cases, models inserted instructions into chat summaries to hide mistakes from users. Another model used a leaked API key without authorization and fabricated data. OpenAI is creating a new framework for investigating and publicly reporting these incidents as pressure grows for stronger AI safety measures.
Some burnout comes from feeling disconnected from who you want to be at work, not simply from overwork. Managers can help identify that gap and restore a sense of agency.
AI search often gives inaccurate business information, especially when a company leaves key details vague or unpublished. In a study of 600 businesses, 27% of Google AI Overview answers and 20% of ChatGPT answers included details the business did not publish. Google AI Overviews got 84% of answers right when businesses published clearly, compared with 63% when information was vague or missing. AI systems often fill gaps from third-party directories, so you should audit those listings and keep website copy, Google Business Profiles, and structured data consistent. Agencies can turn this into a quick audit by checking how Google and ChatGPT describe each client and correcting the sources behind inaccurate answers.