What it’s allowed to do
Permissions, sandboxing and the commands it can run.
Using AI coding agents? Make them safe to run.
AI agents that write code, run commands or touch business data can save a team weeks, and can also do a lot of damage quickly. We review how yours are set up and give your team a safe way to work with them.
What the review covers
How approvals work
Whether they can be skipped, faked or replayed.
Hidden instructions
Whether a web page, a file or another agent could steer it.
Secrets and data
What it can see, and where that goes.
Runaway costs
Budgets, loops and limits.
Logs
What’s recorded, and who can read it.
Data protection
What happens when agents touch personal or client data.
The platform
Whether it’s running in the right place for what it does.
A standards kit for your team
An engineering standard written for AI coding agents: one instruction file and short guides on architecture, secrets, dependencies, deployment, security and testing. Every new project then starts the same, safe way, even when the person building it isn’t a developer.
Work we’ve done
An autonomous coding-agent server, reviewed lens by lens
From permission bypass and sandbox escape to approval replay and data-protection governance. Findings were verified by separate reviewers from the ones that raised them.
A standards kit for a client team
A drop-in framework that lets a non-developer build small apps with an AI coding agent safely, on one agreed stack.
How we keep findings accurate
Evidence for every finding
Each finding is written up with its evidence.
Checked by someone else
Findings are verified by separate reviewers from the ones that raised them.
You choose where the AI runs
Tier 1
A hosted model, through a UK or EU endpoint
The best quality for the least effort. Before any of your data goes to a provider, we check that its business terms rule out training on it, and we tell you which region does the processing.
Tier 2
Inside your own cloud account
The model runs in your AWS or Azure account, under your access controls and your logs.
Tier 3
A private model on hardware in the UK
Open-weight models on your hardware or ours. Nothing goes to an outside AI provider. Our own research and document reading run this way.
Storage location and processing location aren’t the same thing. We’ll tell you both, in writing.
Related pages
Questions
We only use a coding assistant in the editor. Do we need this?
Probably not a full review. The standards kit may still help.
Do you tell us to stop using agents?
Rarely. Usually it’s a handful of changes that make them much safer.
Get in touch
Tell us how your team uses AI agents.
One paragraph is enough. We’ll reply with honest thoughts on whether we can help, and what it would cost.