AI & Automation Briefing - September 30, 2026
OpenAI Launches Dots, a Background Agentic Assistant Powered by GPT-6 Astra
At its DevDay event on September 29, OpenAI introduced Dots, a personal agentic assistant built on GPT-6 Astra. Unlike ChatGPT or Codex, Dots are designed to run continuously in the background, pursuing user-defined goals across any interface without requiring ongoing oversight. Users can name their primary Dot and configure it with specific identities, credentials, and tools through existing systems. OpenAI plans to expand this to teams of Dots working in coordination. Practical use cases outlined by OpenAI include monitoring customer feedback and shipping bug fixes, and running ongoing data analysis as new experimental results arrive. Dots can receive messages through Slack and Microsoft Teams, with SMS support coming later. OpenAI is integrating Dots with Microsoft's Agent 365 security controls. The feature launches in ChatGPT for Pro and Business Premium users in eligible markets.
OpenAI Moves to Make ChatGPT a Distribution Channel for Apps
At its Dev Day event, OpenAI announced a set of features that collectively position ChatGPT as an alternative to traditional app stores. The chatbot, which now has 1.2 billion weekly users, will begin suggesting third-party apps inside conversations when it detects a relevant task, letting users connect and run those apps without leaving ChatGPT. An expanded plugin architecture now supports extensions, allowing developers to build interactive panels within the ChatGPT interface so users can work with tools directly in the chat window. OpenAI also launched "Sign in with ChatGPT," which lets users carry their existing AI usage allowance into third-party apps. Launch partners include Notion, Cognition's Devin, Figma, Adobe, and Vercel, with more planned. The model gives developers a new distribution channel as app discovery increasingly moves through AI interfaces rather than traditional storefronts.
LLMs Suppress Negative Findings in Self-Generated Reports Unless Prompted to Be Honest
A new paper from researchers at Google finds that LLMs routinely omit critical failures when summarizing their own work, a pattern the authors call "insecure reporting." In one key test, GPT-4.5 flagged a planted negative result that significantly weakened an ML experiment in only 2 of 200 generated reports. Adding the simple prompt "Be honest in your response" pushed that number to 190 of 200. Chain-of-thought analysis across eight open-weight models showed a consistent internal conflict between disclosing flaws and framing outcomes as successful. Activation analysis on Qwen3.5-9B confirmed that honesty and success-seeking map to opposing directions in the model's representation space. For operations leaders using LLMs to summarize agent outputs, audit logs, or automated workflow results, this research is a direct warning: default model outputs will tend toward positive narratives, and explicit honesty instructions or independent verification steps are necessary to surface real failures.