AI news · 10 stories

AI news for Nov 17, 2026

Apollo begins embedded AI-safety evaluation pilot at OpenAI · Apple launches study of how people use health AI · OpenAI expands free ChatGPT access for academic researchers

News card: Apollo begins embedded AI-safety evaluation pilot at OpenAI
Apollo Research

Apollo begins embedded AI-safety evaluation pilot at OpenAI

Apollo Research announced November 16 that six evaluators will work inside OpenAI for four months with scoped access. They will examine whether models try to undermine alignment or control mechanisms during development, bringing independent scrutiny closer to training.

Read at Apollo Research
News card: Apple launches study of how people use health AI
Apple

Apple launches study of how people use health AI

Apple announced November 15 a US research study with OpenEvidence examining trust, comprehension, and everyday use of conversational health AI. The study can enroll up to 200,000 eligible adults and aims to inform responsible design of health-AI experiences.

Read at Apple
A glowing antenna broadcasting neon signal rings
OpenAI

OpenAI expands free ChatGPT access for academic researchers

OpenAI announced November 16 an expansion targeting up to 200,000 researchers worldwide. Applications for 50,000 new grants open November 23, with each grant providing ChatGPT Plus to a recipient and up to two collaborators, widening access to AI research tools.

Read at OpenAI
News card: OpenAI advances computer-use agents for healthcare workflows
OpenAI

OpenAI advances computer-use agents for healthcare workflows

OpenAI announced November 16 computer-use capabilities for healthcare through its API, ChatGPT for Healthcare, and ChatGPT for Clinicians. Work with organizations including CVS Health and Cigna targets tasks across disconnected software, with approval controls and audit logs supporting oversight.

Read at OpenAI
News card: Google rolls out 66 regional voices in Gemini
Google

Google rolls out 66 regional voices in Gemini

Google began rolling out 66 regional voices in Gemini on Android and iOS on November 16. The selected voice works across Gemini Live and spoken chat responses, bringing locally familiar cadence and inflection to conversations.

Read at Google
News card: MLCommons releases MLPerf Training v6.1 results
MLCommons

MLCommons releases MLPerf Training v6.1 results

MLCommons published MLPerf Training v6.1 results on November 16 with submissions from 24 organizations. New recommendation and large-model post-training workloads broaden the suite’s coverage and provide comparable evidence for evaluating AI training systems.

Read at MLCommons
News card: Microsoft details TokenEconomizer for more efficient coding agents
Microsoft Research

Microsoft details TokenEconomizer for more efficient coding agents

Microsoft Research described TokenEconomizer on November 16, an open-source plugin that learns from prior coding-agent sessions to reduce repeated work. Across 277 bug-fixing tasks, its evaluation reported 18.4–30.9% lower costs with broadly stable task success.

Read at Microsoft Research
News card: Google reports autonomous patch-to-exploit framework in the wild
Google Cloud

Google reports autonomous patch-to-exploit framework in the wild

Google’s November 16 threat report describes exploit_agent, a framework it found comparing software changes, drafting exploits, and executing them without human intervention. The finding highlights the shrinking window defenders have to respond after patches disclose vulnerabilities.

Read at Google Cloud
News card: The King’s Trust and OpenAI launch AI-skills partnership
OpenAI

The King’s Trust and OpenAI launch AI-skills partnership

The King’s Trust and OpenAI announced a £1 million partnership on November 16 to support up to 25,000 young people in the UK over 18 months. The program prioritizes those outside education, employment, or training and aims to build practical AI skills for work.

Read at OpenAI
News card: Amazon Bedrock adds explicit caching for open-weight models
Amazon Web Services

Amazon Bedrock adds explicit caching for open-weight models

AWS announced November 16 that open-weight models on Amazon Bedrock now support explicit prompt caching alongside model-dependent implicit caching. Developers can mark reusable prompt prefixes through the Chat Completions API to reduce repeated processing, input costs, and latency.

Read at Amazon Web Services

Get the next episode first

Shorts and deep dives on AI system design land on @AI.JoinDev first. Subscribe, or follow the RSS feed.