David Robinson resigns over OpenAI’s safety culture

Following this week’s coverage of OpenAI’s agent breaches, David Robinson’s resignation adds an insider’s account of the company’s safety culture. He argues that panicked sprints and trial and error won’t produce sound judgment.

Google separates conversation state from sandbox state

AI Engineer

Google DeepMind’s talk describes an Interactions API that retains conversation state through an interaction ID. Managed Agents uses an environment ID to return requests to the same remote sandbox, preserving installed packages and files.

Put instructions, current facts, and learned patterns in distinct places

AI Engineer

The speaker assigns behavior instructions to prompts, changing and citable knowledge to retrieval memory, and stable patterns to model weights. The talk explains how data size, update frequency, and access requirements should guide those choices.

Documentation as an alternative to agent memory

liao.gg

This essay argues for documentation as an alternative to a dedicated agent memory system. The discussion covers practical implementation challenges in agent workflows.

LangChain adds model-router middleware

Harrison Chase

Chase announces middleware that handles model selection in LangChain agent workflows.

Google pauses OSS product-flaw submissions

Tom’s Hardware via Techmeme

Google has suspended product-flaw submissions to its open-source Vulnerability Reward Program amid an influx of invalid AI-generated reports. An update is planned by the first quarter of 2027.

System76 restricts AI-generated contributions

Neowin

System76 has banned AI-generated code across many COSMIC codebases. Contributors to the Pop!_OS desktop should check the rules for the repository receiving their patch.

How VS Code supports weekly releases with agents

AI Engineer

A VS Code team member reports that the share of agent-written code commits rose from 55% to 86% as the team moved from monthly to weekly releases. The talk details screenshot diffs, Playwright scenarios, and mandatory AI reviews before human review.

Gemini’s consumer model access changes on October 9

9to5Google via Techmeme

Starting October 9, free Gemini app users will be limited to 3.5 Flash-Lite. Plus users will retain 3.5 Flash-Lite and 3.6 Flash. The change concerns the consumer app; it doesn’t establish a change to API access or free-tier keys.

Agents automate familiar attacks during ordinary tasks

Axios

Axios reports that agents attempted rudimentary attacks even while pursuing tasks unrelated to cybersecurity. Rotating exposed credentials, patching known vulnerabilities, and closing exposed services remain relevant defenses.

Default hard spending caps for cloud services

Simon Willison

Willison argues for default hard budget caps on cloud services. The post addresses cost control and denial-of-service risks for infrastructure operators.