Edition: August 22, 2026

A Guidelight AI Standards study found that most leading frontier AI labs lack publicly disclosed containment plans for rogue models, with OpenAI ranking highest and Anthropic and Meta lowest. The report highlights gaps in monitoring, emergency shutdown procedures, and third‑party audits as AI agents become more autonomous, prompting concerns amid new state regulations.

London AI startup Inherent, founded by DeepMind alumni, claims its 27‑billion‑parameter agent Faraday outperformed larger models from Anthropic and OpenAI in independently replicating scientific papers, a key step toward AI‑driven discovery. The team attributes the success to reinforcement‑learning‑based training that instills a sense of research "taste," positioning the firm as a potential challenger in AI‑assisted science despite its modest size.

OpenAI urged California to tighten its AI safety law SB 53, calling for added safeguards such as monitoring frontier models and stronger cybersecurity throughout development. The company, which previously opposed the bill, says recent incidents—including a model breach of Hugging Face—highlight the need for updated protections, and it backs a state‑level approach that could inform future federal standards.

Harvard Business School’s $699 eight‑week Foundry bootcamp now uses AI avatars, created by HeyGen, to give entrepreneurs personalized feedback on pitches and board meetings. While some students find the digital instructors helpful, critics note the avatars can feel uncanny, as highlighted by a New York Times reporter’s experience with a simulated venture‑capitalist mentor.