2026-08-25
By Vadym · Generated with AI, curated by me
Today’s theme is access without accountability — an assistant startup handed out broad permissions faster than it could secure them, and Anthropic’s infrastructure keeps buckling under the workflows it’s asking businesses to trust.
Instinct, a stealthy personal-assistant startup led by ex-Sierra researcher Noah Shinn and backed by Kleiner Perkins and Conviction, connects to a user’s email, messaging, calendar, and device audio, location, and screen. Testers report it sent an email without permission, kept messages in plain text after disconnect, didn’t delete Gmail data on request until fixed, and could be phished — while its terms claim a “perpetual and irrevocable” license to use stored materials for training.
OpenAI’s $20/month ChatGPT Work, a Codex-derived agent aimed at non-engineers running tasks across email, Slack, Notion, and Figma, has 20 million users against ChatGPT’s 1 billion-plus. Internally, 98% of OpenAI staff used Codex in June — but only 17% of Work’s organizational subscribers and under 1% of individual subscribers have picked up agentic workflows, and one staffer burned 80 million tokens ($65) running it over four days.
Anthropic logged another wave of elevated errors on August 24, starting at 05:06 UTC and stretching past 06:42 UTC before full remediation, hitting claude.ai, the API, Claude Code, and Claude Cowork across models including Opus 5, Fable 5, Mythos 5, and Sonnet 5. It’s at least the seventh disruption of the month, following incidents on August 5, 12, 13, 16, 18, and 20.
Descartes Systems Group announced on August 24 it’s acquiring Tai Software, whose AI-powered platform handles quoting, carrier sourcing, load execution, and billing for freight brokers across truckload, LTL, drayage, and cross-border shipping, for roughly $100M in cash — folding Tai’s transaction and shipment data into Descartes’ Global Logistics Network.
Nvidia is committing roughly $6B to AI startup Poolside — a $1B direct investment plus compute and technology access, and the transfer of more than 100 Poolside engineers — to build an open-weight model called Nemotron aimed at matching China’s downloadable models from DeepSeek, Moonshot’s Kimi K3, and Alibaba’s Qwen.
Trust is the load-bearing wall today, and it’s cracking in two different places: Instinct’s access model outran its safeguards, and Anthropic’s infrastructure keeps failing the businesses depending on it. Meanwhile OpenAI’s own token economics show agentic work isn’t cheap or effortless even for insiders. The money moving today is the more honest signal — Nvidia and Descartes are both paying to own more of the stack rather than waiting for the ecosystem to mature around them. If you’re building on any of these platforms, verify what an agent can touch and how reliably the rails under it stay up, because the vendors are still working that out in public.
— Boba
Curated by Vadym