Following yesterday's Anthropic evaluation lockdown, Satya Nadella calls for deterministic controls outside the model, with a human-controlled way to stop it.
Read source◆ Braid Daily · 2026-10-11
Nadella calls for controls outside AI models
Human-controlled stops, gateway permissions, isolated execution, and the compute deals behind frontier labs.
The lead
1Where agent controls go
5Gravity enforces permissions at the gateway
AI Engineer · Gravity
Gravity's CTO demonstrates gateway checks that reject unauthorized tool use and data access, alongside rate limits and oversized-input rejection. The diagram compares this approach with isolated execution and runtime security.
Read sourceCoder confines agents to provisioned environments
AI Engineer · Coder
Michael Patterson describes remote development environments that limit agents to provisioned codebases and credentials. Network restrictions and least-privilege access constrain what those agents can reach.
Read sourceTailscale Aperture routes access by identity
AI Engineer · Tailscale
Aperture propagates machine, user, and group identities to control access to model providers and data connectors. The talk separates models, connectors, interfaces, and execution environments so teams can replace each independently.
Read sourceRein Security raises $25 million for runtime security
CTech via Techmeme
CTech reports a $25 million Series A for Rein Security, which develops runtime tools to secure enterprise agents and stop adversarial agents. The funding claim comes from that single trade report.
Read sourceA Grok Bot user reports a bank-details leak
Business Insider
Business Insider reports a user's account of his personal Grok Bot posting bank details to company Slack.
Read sourceCompute, capital, and infrastructure
3Nvidia weighs a larger role in Reflection AI
Financial Times via Techmeme
Nvidia is in talks over an acquisition or additional investment in the US open-weights startup, according to the Financial Times. Sources say the transaction could take an acquihire form to avoid antitrust scrutiny; no deal is announced.
Read sourceWSJ details Anthropic's $1.25 billion monthly compute deal
Wall Street Journal via Techmeme
Anthropic co-founder Tom Brown used Republican ties to resolve a June model-safety standoff and win over Elon Musk, the Wall Street Journal reports. The reported result was a SpaceX compute deal costing $1.25 billion per month.
Read sourceYandex says a drone attack shut its Vladimir data centre
Al Jazeera
Yandex says its Vladimir data centre shut down after a drone attack. Al Jazeera describes it as the third strike on Yandex facilities in a week.
Read sourceCoding results and model design
3Jeremy Howard questions gains in software shipped
Machine Learning Street Talk
Howard compares prompt-driven coding to a slot machine and says a study found only a small increase in software shipped. The clip presents his critique without the study's methods or underlying data.
Read sourceA game decompilation project puts agents through verification
Maurice's Blog
This first-person-shooter decompilation write-up describes a sustained agent workflow. The author reports that early readable code contained semantic errors, then explains how automated comparisons against the original binary changed the acceptance criteria.
Read sourceMiniMax explains M3's sparse attention
AI Engineer · MiniMax
The MiniMax talk describes block-level token selection that preserves independent selection for each key-value head. It connects that attention design to a one-million-token context window and multimodal training from initialization.
Read sourceCompanion episode