AI Daily Dev — June 19, 2026

OpenAI o3 Cracks 18 Stumped Rare Disease Cases in Boston Children's NEJM AI Study

Boston Children's, Harvard, and OpenAI ran o3 Deep Research over 376 previously unsolved pediatric genomes and got new diagnoses on 18 (4.8% extra yield).
Breakdown: 10 neurodevelopmental, 4 neuromuscular, 2 early childhood psychosis, 2 sudden unexpected pediatric deaths — all confirmed by clinicians.
Hospital says the program has unlocked 40+ previously unsolved cases, saved ~60,000 hours of analyst time, and redeployed >$7M in labor costs.
Genetic investigations director Catherine Brownstein: 'a total game changer.' Published June 18 in NEJM AI.

research openai.com

Admins now provision MCP connectors organization-wide through Okta; users get connector access on first login with no per-user OAuth dance.
First implementation of MCP's new Enterprise-Managed Authorization extension, built on Okta's Cross App Access (XAA) standard.
Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase live at launch; Slack coming soon.
Hubspot, Ramp, and Webflow are rolling it out across their teams; VS Code adds matching IDE support.

tools claude.com

OpenAI says GPT-5.5 Instant now matches its frontier Thinking models on the hardest health evaluations, including HealthBench Professional.
230M+ people ask ChatGPT health questions weekly; OpenAI reports a 71% drop in factuality issues vs GPT-5.3 Instant.
Evaluation is steered by OpenAI's Global Physician Network — 262 doctors, 26 specialties, 59 countries.
Biggest health-capability jump since GPT-5; pitched as the consumer-side counterpart to the Boston Children's o3 study.

models openai.com

New plan defends Google's internal systems against potentially adversarial AI agents using the same playbook as a rogue employee with office keys.
Companion paper 'Three Layers of Agent Security' splits the problem across single agents, multi-agent systems, and ecosystem-wide cyber defense.
Builds on June 11's $10M multi-agent safety funding call with Schmidt Sciences, ARIA, Cooperative AI Foundation, and Google.org.
MIT Tech Review framed the work as DeepMind's first concrete take on what happens when millions of agents start interacting.

research deepmind.google

xAI's flagship Grok models now run inside Databricks Agent Bricks, sitting next to OpenAI, Anthropic, and Meta models on the same governed platform.
Agents can reason over Lakehouse data directly — no external pipelines, no copying enterprise data out of Databricks' governance layer.
Zero data retention endpoints on the xAI side; Databricks doesn't train foundation models on customer data.
Announced June 18 at the Databricks Data + AI Summit; Musk shared the integration on X.

tools x.ai

Still no restoration date; Commerce talks continue with no announced deal one week into the export-control suspension.
Refund window for users who upgraded June 9–14 closes tomorrow, June 20 — last business day to claw back a charge for a model nobody can use.
Polymarket 'restored before July 1' has slid from ~80% to 58%; 'by July 10' sits around 74%.
Anthropic still calls the underlying jailbreak 'not serious'; the open-source camp keeps pointing at this week's MIT-licensed GLM-5.2 as the alternative.

industry anthropic.com