See what Copilot would hand an attacker.
Microsoft 365 Copilot answers with whatever the person asking can open. We map what that is, then test what an attacker, a careless prompt or a poisoned document can make it do, under rules of engagement you sign.
Copilot inherits every permission nobody cleaned up.
Copilot works with the permissions of the person asking. It will only retrieve what that user could already open, and that is often far more than anyone has checked. The UK government’s own cross-government Copilot experiment, which gave 20,000 civil servants a licence in late 2024, warned that Copilot’s ability to search with a user’s access “may highlight issues when users have access to files they shouldn’t”.
Readiness scans check the settings. They do not test what an attacker, a careless prompt or a poisoned document can make Copilot do. The NCSC’s view is that prompt injection may never be fully mitigated, because a language model enforces no security boundary between the instructions it is given and the content it reads. The work is to reduce the likelihood and the impact, and to prove that you have.
Sources: Microsoft 365 Copilot Experiment: Cross-Government Findings Report, GOV.UK, 2 June 2025, section 2.1. NCSC, Prompt injection is not SQL injection (it may be worse), 8 December 2025. Both read 17 November 2026.
Two weeks, read-only, and you know what Copilot can reach.
Request a bespoke quote →Review it, attack it, keep it tested.
Each has a fixed scope and a fixed length, agreed with you before anything starts, and each is quoted to your organisation.
Copilot Readiness and Exposure Review
Two weeks
Copilot Red/Blue Exercise
Around six weeks
Copilot Continuous Assurance
Monthly
Rules of engagement first. Your team stays in control.
- 01
Agree the rules of engagement
The scenarios, the pilot group, the timings and the stop conditions, written down and signed before any testing starts.
- 02
Run the scenarios
Each scenario runs against the pilot group only, inside the signed scope. Your team can pause or stop any of them.
- 03
Defend and tune
Your team defends as each scenario runs. Where one gets through, we work on the detection with your team and re-run the scenario to check the change holds.
- 04
Leave it with your team
The findings report, the tested scenario library, the monitoring and change playbook, and a training day, so your team can run the scenarios again without us.
Nothing is tested until the rules of engagement are signed. Your team keeps control of your systems throughout, and nothing outside the signed scope is touched.
Copilot assurance at Velinor is led by the two founders. James A. Lang leads on Secure by Design and AI governance. He retired a Warrant Officer Class 1 after 22 years in the Royal Signals and was appointed the MOD’s first AI Cybersecurity Product Manager. Ben Brand is a former British Army intelligence analyst who ran the MOD’s Vulnerability Management Support Team.
Through the AI DPS, or direct.
Velinor is an appointed supplier on the Government Commercial Agency’s Artificial Intelligence Dynamic Purchasing System, RM6200, so public sector buyers can award work to us through it. Regulated firms, and anyone who prefers to, can contract with us directly.
What buyers ask first.
- 01
Do you access our tenant?
For the review, only with read-only access that you grant, and we change nothing. For an exercise, we work inside rules of engagement you sign before any testing starts, against the pilot group you name, and your team keeps control of your systems throughout.
- 02
How long does it take?
The Readiness and Exposure Review takes two weeks. The Red/Blue Exercise runs for around six weeks. Continuous Assurance runs monthly, with a review every quarter.
- 03
What do you need from us?
A named technical contact and read-only access for the review. For an exercise, a pilot group, time from your security team while the scenarios run, and a signature on the rules of engagement before we start.