Apple locks down agent access

Apple is tightening who gets the keys to a Mac while Anthropic and Chatham focus on the people and checks behind useful AI.

Today’s issue covers agent permissions, a $100 million training bet, and trade checks cut from 30 minutes to four.

AI agents have started asking for one of the broadest permissions a Mac can grant: access to files, mail, messages and browsing history.

Apple says some developers are using Full Disk Access in ways that expose that data without users fully understanding what they approved. The same permission can also reveal private messages sent by other people.

The company will add new controls that require a more explicit action before an app gets that level of access. Apple has not given a release date or described the new approval flow yet.

The warning is unusually direct. Apple says the risk will grow "substantially" as agents become more capable and autonomous.

That changes the default for Mac software that wants to act on a user’s behalf. Broad permission requests will face more friction.

Products built around silent, always-on access may need to narrow what they collect or explain it far better.

Founders context: Treat an internet-enabled agent like a junior employee with a browser. Give it only the files and accounts the job needs, log what it does, and make broad permissions expire.

Anthropic is spending $100 million to train 10,000 engineers who can take Claude from a demo to a working system inside a large organisation.

The first cohorts include engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. The target is 10,000 graduates by the end of 2027.

This is closer to a residency than an online course:

The first cohorts are already running in San Francisco, New York and London. Entry is through an Anthropic account or partner manager, not open enrolment.

The programme says something blunt about enterprise AI. Better models are not enough when companies lack people who can choose a useful problem, clear security review and get the result into daily work.

Founders context: Before buying more AI licences, name one person who owns the first live project from process mapping through review and adoption. A tool without that owner usually stays a demo.

AI IMPLEMENTATION

This is how Chatham Financial cut a routine trade check from about 30 minutes to under four without handing the decision to a model.

The capital-markets adviser first mapped the minimum evidence needed to confirm that a system record matched what a client authorised and what traders executed.

It then built an application with Codex that gathers the evidence, compares key terms and flags mismatches for a person to review.

The important part sits around the software:

Chatham routes simpler analysis and non-production testing to cheaper models. It reserves stronger models for work where accuracy is worth the extra cost.

The proposal is portable: pick a repetitive check with a clear evidence trail, let software assemble and compare the material, then keep a specialist on exceptions. The time saving comes before full automation.

🔎 Cloudflare adds web search to AI Gateway. Agents can now query Ceramic, Exa or Linkup through one API, with search calls appearing in the same logs and billing controls as model traffic.

🖥 NVIDIA doubles the memory in DGX Spark. The new 64GB desktop system is aimed at running larger local models and moving work between a personal machine and NVIDIA cloud infrastructure.

🏘 Amazon courts towns around its data centres. The company is offering local funds and planning support as communities push back on the land, water and power demands behind AI infrastructure.

🧭 Cloudflare releases open decision models. Clef and Clef Flash are built to choose between options rather than produce long answers, with an open fine-tuning route for specialised decisions.