
Lab
Tools I wrote to practise what I preach.
Tools that are the method, running. The largest, fleet-watch, has its own page.
fleet-watch
One console for every server you need watched: break-in detection, vulnerability and update checks, signed updates rolled out in stages, a local AI first read and a hash-chained evidence log.Hash-chained logEach record carries a fingerprint of the one before it. Change or delete any record and every later fingerprint stops matching, so tampering is visible.Learn more → I run it on the servers I am responsible for, clients' included, and offer it as a managed service. See fleet-watch.
The build-watcher
A small tool that watches the upstream builds of the software I depend on. For each new build it works out what the build fixes, and a model running locally on my Mac classifies it. Security fixes raise an alert.
It is what woke me at 01:55Why 01:55The time, in the night after the PHP2Shell release, when the corrected PHP images were built and the alert went off.Learn more → on the PHP2Shell night, when the patched PHP images were built. It is now moving into fleet-watch. The story is in Three weeks in October.
Compliance cockpit
A web app I built at a B2B go-to-market AI platform to work through SOC 2 and ISO 27001 tasks. It exists because vulnerabilities now arrive faster, helped by AI-assisted discovery, and triage and evidence have to keep pace.
- One view over the compliance platform's tasks, per control and per owner.
- An AI-assisted session for each task, to draft evidence that a person then checks.
- Evidence written to documents, with PDF export for the auditor.
- Kept off the public internet, behind an identity-aware access gate.
The same thinking applies to compliance readiness for your organisation.
Engineering agents, on a short lead
Small agents for my own engineering work: they watch pull requests and production monitors and draft status updates. A person decides what ships. They sit next to skills, hooks and project rules that keep AI-assisted code reviewable, and I benchmark local models for work where data must stay in-house.
An agent only does useful work with real access, so the access is narrow, short-lived, logged and approved by a person:
Egress through an allow-list
Agents reach the network only through a proxy with a list of permitted destinations. Everything else is refused and logged.
SSH certificates per task
Access to a server is a certificate minted for one task, restricted to one source address and short-lived.
SQL through a controlled integration
Database queries go through a controlled integration with its own limits. The agent never holds a database password.
A secrets runner
A command that needs a secret gets it injected at run time. The value is scrubbed from everything the agent can read.
A guard the agent cannot edit
The rules live in a guard owned by root, outside the agent's reach. It cannot rewrite its own limits.
A person approves
Risky actions stop at a deterministic approval step: the same request always asks the same question, and a person answers it.
Why I went this far is on the Security page. Setting it up for a team is part of AI-assisted engineering.