Start Here

Begin with Synoptikon

A short guide to Synoptikon, organised around Big Reads, Explainers, Perspectives, and Workbench.

What is Prompting Trust?

The weekly companion to Synoptikon, for current developments and shorter notes.

Feeds

Follow Synoptikon by feed, email, or curated sources without relying on the homepage.

Can you still trust your PC?

A simple checklist for testing whether a local AI agent has a real control boundary.

Related Big Read: The localhost illusion: why the trust boundary has moved inside the machine

Related assets: Agent identity as the control boundary; Agentic blast radius; AI supply chain verification

Localhost used to feel reassuring.

If a service was bound to the local machine, many teams treated it as a lower-risk surface. It was close to the user, hidden from the internet, and outside the normal perimeter story.

That assumption starts to break down when local AI agents gain access to tools, the browser, files, and a shell, while extensions, skills, and workflow chains add new paths into the local environment.

The question is no longer where the agent runs.

The question is what it can do.

Why location is a weak control

Local interfaces can still be reached through browsers, extensions, misconfigurations, exposed ports, and workflow chains. A local model host or desktop agent can become a control surface if it accepts instructions, calls tools, or handles sensitive context without strong authentication.

The risk is not only remote compromise.

It is over-trust by design.

If the agent inherits the user’s full authority, accepts ambient local requests, and lacks a clear revocation path, locality has become a comfort story rather than a control.

This simple toolkit helps you test whether a local AI agent has a real control boundary. It is deliberately high-level. Detailed mTLS configuration, local-agent hardening, and incident procedures can be developed as separate assets where the use case justifies them.

Local agent checklist

Use these five control areas before a local AI agent receives access to sensitive systems, data, or workflows.

01

Local interface trust

02

Authentication and transport

03

Semantic supply chain

04

Non-human identity

05

Culture and shipping discipline

The executive test

Ask your engineering lead:

Sign-off rule

If this agent received a hostile instruction through an ordinary local interaction, what could it reach before you stopped it?

If your answer is unclear, the boundary is not localhost.

The boundary is missing.

Previous Post
The localhost illusion: why the trust boundary has moved inside the machine - featured image

The localhost illusion and why the trust boundary has moved inside the machine

Next Post
The dependency nobody vetted: how AI tooling became a supply chain vulnerability - featured image

AI tooling is becoming a supply chain vulnerability

Subscribe to Prompting Trust

Subscribe to Prompting Trust to receive The Weekly Context.

Prompting Trust is the newsletter layer connected to Synoptikon. It carries current developments, useful links, and shorter notes, while Synoptikon holds the longer arguments and working library.

Learn more about Prompting Trust.