AI News from August 21, 2026

Rillet reportedly raised $100 million and reached unicorn status in 48 hours, putting AI accounting startups under sharper investor and buyer scrutiny.

Nvidia researchers report that memory controls and a supervisory agent helped Claude Opus 5 reach a perfect ARC-AGI-3 score, reshaping agent design.

Nvidia’s latest deals are being framed as a route beyond AI-chip sales, but limited source evidence leaves the counterparties and terms unconfirmed.

NVIDIA’s new agent-stack guidance puts final security authority in runtimes and infrastructure, not editable harnesses, as AI agents gain autonomy.

NVIDIA says its AVO agent reached a perfect ARC-AGI-3 score, highlighting how memory, supervision and tools can extend model performance.

NTT’s “AI coworkers” framing and Burges Salmon’s EU AI Act update point to a shift from copilots to governed, accountable workplace agents.

Nvidia is reportedly discussing a potential deal with Korean AI chip startup Rebellions, signaling broader interest in alternatives for AI inference and data-center workloads.

Anchorage’s CEO says autonomous AI agents may need bank accounts, sharpening questions about payments, identity, controls, accountability, and trust.

AWS has detailed a no-code workflow linking Snowflake, SageMaker Canvas and Amazon Quick to help business teams build fraud models and dashboards.

Binance is reported to have launched Agent OS for AI-driven market access and crypto trading, raising new questions about controls and accountability.

Harvey has introduced Tenet, its first post-trained AI model for legal work, signaling a push toward domain-specific systems built for lawyers.

OpenAI is previewing GPT-5.6 Sol, but limited public evidence leaves builders waiting for verified details on capabilities, access, pricing, and safety.