AI Sector Daily Digest — July 15, 2026

The 24-hour signal

Five stories

1. The U.S. is setting up an AI cybersecurity coordination group

• The White House says it will bring AI developers and essential-services providers together to share vulnerabilities found by advanced AI systems and coordinate responses. 1
• Treasury, the National Cyber Director's Office, the Department of Defense, and the NSA are involved; open-source developers are included, but named company participants were not disclosed. 1
• The mechanism fulfills a June presidential order, but the announcement gave no operating timeline or public member list. 1

2. Reflection AI signs a $1B compute deal with Nebius

• Reflection AI signed a $1B agreement to access compute from Nebius, including Nvidia's latest chips, TechCrunch reports. 2
• The parties did not disclose the amount of capacity or the contract term, so the headline value cannot yet be translated into usable GPU-hours. 2
• For open-model builders, the practical question is moving from whether compute is available to how much guaranteed capacity a deal actually delivers. 2

3. OpenAI's first hardware product is reportedly a screen-free speaker

• Bloomberg reports that OpenAI is developing a portable, screen-free smart speaker as its first consumer hardware product; Reuters carried the report. 3
• The device is described as having cameras and sensors, with planned support for smart-home control, media playback, questions, messages, and ChatGPT capabilities. 3
• OpenAI did not immediately respond to Reuters, and no release date was reported, so this is a product report rather than a launch announcement. 3

4. Chai Discovery raises $400M at a reported $3.8B valuation

• Chai Discovery announced a $400M Series C; Sapphire Ventures confirmed its participation, while SiliconANGLE reports a $3.8B valuation and Index Ventures as lead investor. 4 5
• The company develops models that predict and reprogram molecular interactions, with antibody design among the stated applications. 5
• The commercial signal is pharma adoption rather than an approved medicine: coverage cites Pfizer, Eli Lilly, and Novartis relationships while noting that no AI-discovered drug has yet been approved. 5

5. PM-Bench tests whether agents remember what they promised to do

• A July 14 arXiv paper introduces PM-Bench, a text benchmark that puts agents through a simulated seven-day week of delayed intentions, updates, cancellations, and hidden-state changes. 6
• Across eight model backbones and eight agent configurations, the best setup reached 65.1% macro set-F1; no setup exceeded 50% on cross-day or update-sensitive tasks. 6
• The benchmark is synthetic, but its practical message is clear: agent evaluation needs to measure monitoring, reminders, and state updates, not only final-answer accuracy. 6