AI Daily — September 18, 2026
2026-09-18
openai.com
Introducing Astra for Law
OpenAI has launched a dedicated legal AI platform called Astra for Law, offering frontier models integrated with custom firm workflows, connected legal data sources, and confidentiality controls designed for attorney-client privilege requirements. The product targets law firms directly with legal-grade compliance and security controls, signaling a serious vertical push beyond generic ChatGPT enterprise tiers. This follows the Cooley GO Public case study, which demonstrated ChatGPT being used to surface IPO issues earlier in the diligence process.
www.anthropic.com
Introducing the Life Sciences Verification Program
Anthropic has announced a Life Sciences Verification Program, creating a formal credentialing process for researchers and organizations accessing Claude's capabilities in biomedical and pharmaceutical contexts. The program appears aimed at tiering access controls based on verified institutional affiliation, consistent with Anthropic's ongoing policy focus on biosecurity risks from AI-assisted research. This is a notable structural safety measure that could set a precedent for access gating in high-risk scientific domains.
rss.arxiv.org
Radio-Frequency Convolutional Neural Networks
Researchers propose RF-CNNs, a method that repurposes the frequency mixer hardware already present in wireless radios to perform CNN inference natively in the frequency domain, eliminating the need for dedicated edge AI accelerators. Since frequency mixers inherently multiply signals — equivalent to convolution — existing communication silicon can perform neural network forward passes with no additional hardware, directly addressing SWaP-C constraints on edge devices. If validated at scale, this approach could substantially reduce the cost and power overhead of deploying CNNs on smartphones, wearables, and drones.
rss.arxiv.org
Open ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings
The authors release SonoCorpus, a large open dataset unifying 456,963 ultrasound images and 1.6M expert segmentation masks from 53 public datasets across 24 clinical applications and 17 countries, alongside SonoBase, a foundation model for interactive ultrasound segmentation trained on it. SonoBase outperforms SAM2, MedSAM2, and MedSAM3 across 15 evaluation datasets spanning new organs, devices, and geographies, and matches specialist models trained per-dataset — a meaningful generalization result for medical imaging AI. The open release of both dataset and model addresses a key fragmentation problem in clinical ultrasound AI, where narrow single-task models frequently fail under distribution shift.