Who let the agents in

New methods are emerging for AI agents to access websites requiring logins, such as ChatGPT Work’s secure sign-in flow that avoids direct chat exposure of credentials. The article also highlights recent developments in AI models like GLM-5.3-Flash and Gemini 3.5 Transcribe, and a new OpenAI chip outperforming Nvidia’s. Additionally, it touches on memory management in AI, investigative reports, and various AI agent tools and platforms.

• ChatGPT Work introduces a secure sign-in flow for agents to access websites without sharing credentials in the chat.
• Updates include new AI models: GLM-5.3-Flash, Muse Image in Meta’s API, and Google’s Gemini 3.5 Transcribe.
• OpenAI’s new chip ‘Jalapeño’ reportedly outperforms Nvidia’s Blackwell chips in speed and efficiency.
• Claude Chat and Cowork now share memory, raising concerns about unintended information transfer.
• Several new AI agent tools and platforms are mentioned, including Gravitee for accountability and opencomputer.dev for deploying agents.
• Scheduled tasks in ChatGPT can now be triggered by events like Slack messages or new emails.
  https://bender.layer3.press/articles/d2c170b0-9634-4a58-bdcf-cc923d6f0fa4