Apple plans tighter Mac access controls for AI agents

Apple says more autonomous agents make broad access to files, mail, messages, and browsing history riskier. Its planned macOS controls will require more explicit user action before granting Full Disk Access; the announcement gives no rollout date or detailed mechanics.

Docker puts coding agents inside microVMs

AI Engineer / Docker

Docker’s SBX demonstration gives each agent session its own kernel and filesystem, with configurable network rules and read-only mounts for related repositories.

OpenAI’s breach-review bill

The Guardian

Following yesterday’s breach coverage, OpenAI tells the Guardian its 50-petabyte review costs over $500,000 a day.

Georgia confronts ballot privacy risks

The Guardian

Georgia held an emergency meeting after research showed AI could link public election records to voters’ ballots.

A judge rules a Flock search unconstitutional

404 Media via Techmeme

404 Media reports that a judge found an officer’s warrantless license-plate search through Flock violated the Fourth Amendment. The ruling concerns that officer’s use of the nationwide surveillance network.

Broadcom’s syndicate seeks $60 billion for AI chips

Bloomberg via Techmeme

Following yesterday’s Broadcom financing item, Bloomberg reports that its Wall Street syndicate is gathering $60 billion for Anthropic and other companies. The financing is still being assembled; the deal isn’t complete.

Amazon ends county-official NDAs for data centers

Wired via Techmeme

Amazon says it has stopped using nondisclosure agreements with county officials on data center projects. AWS chief Matt Garman acknowledges that community opposition is leading to moratoriums.

Data center opposition spreads beyond the US

CNBC

CNBC reports on communities across Europe and Asia pushing back against the rising costs of AI infrastructure.

Kolibri-1 releases Apache-licensed open weights

Aleph Alpha / Hugging Face

Aleph Alpha’s mixture-of-experts model has 78 billion total parameters, with 3.46 billion active per token. Its Apache-2.0 weights support a one-million-token context limit; the model card recommends shorter contexts for complex tasks and efficient serving.

The Redis creator’s ds4 runs models locally

DwarfStar / HN

DwarfStar 4, from the creator of Redis, runs open-weight language models on your own hardware.

Browserbase measures reliability per transaction

AI Engineer / Browserbase

In Browserbase’s example, each step succeeds 99% of the time. Across a 100-step workflow, that yields roughly 36% success. The talk shows how dedicated tools and deterministic checks reduce the number of steps delegated to a model.

Anthropic commits $100 million to deployment training

Anthropic via Techmeme

Claude Frontier Academy aims to train 10,000 Frontier Deployed Engineers with a $100 million commitment. Accenture and Bain are among the first participating organizations.