- A live assistant nobody has tested
- Your brief and its limits
- Running the assistant locally
Red-team and harden a council’s AI benefits assistant
Brightmere Council has a retrieval assistant answering housing benefit questions, and it will tell you anything if you ask it in the right order. Break it methodically against the OWASP LLM Top 10, then build the defences and the regression suite that stop the same attacks working tomorrow — with the evidence a public-sector assurance panel needs.
PCREATED BYPrepmeetsProject studio, Prepmeets
THE CHALLENGE → YOUR OUTCOME
Real context. Work that matters.
Brightmere’s assistant went live after a four-week pilot and has never been security tested. You run a structured adversarial assessment against it, implement the defences, prove them with a regression suite that runs in CI, and write the report an assurance panel can act on — including the residual risks you could not close.
What’s at stake
Your responsibility: Work as the ai security engineer—investigate the requirements, implement your approach and verify the result.The assistant was built in a four-week pilot to take pressure off a benefits phone line with a 20-minute wait. It answers questions about eligibility, evidence and timescales from the council’s published guidance, and it can look up the status of a claim if the resident gives their reference number. The pilot was judged a success: call volume fell 23% and resident feedback was good.
Success means: A findings table mapped to the OWASP LLM Top 10 with severity argued from resident impact, defences implemented and measured, a regression suite in CI, a false-positive measurement showing the assistant still serves residents, and an honest residual-risk section. Explain the decisions and checks behind your implementation.
WHAT YOU’LL BUILD
01LLM security
02Prompt injection
03OWASP LLM Top 10
Brightmere Council housing benefit assistant
INCLUDED WITH YOUR PROJECT
A workspace, not just a document.
Your brief, a Jira-style task board and an interview studio—all in one place.
A look inside your workspace
Start with the full project brief
Build step by step Track your progress Own your interview story
Know where you’re heading.
- 01A live assistant nobody has tested
- 02Your brief and its limits
- 03Running the assistant locally
- 04Four trust boundaries
- 05Baseline, attack, defend, prove
- 06Regression, false positives, model swap
- 07What the panel receives
- 08Submission, legality and cleanup
Practise explaining your decisions
Which decision mattered most in your solution, and how did you verify it?
Use your own implementation to connect Situation, Task, Action and Result into one interview story.
Before you begin · tools & running costs
Prerequisites: Your own accounts and access to promptfoo, garak, Llama Guard, Chroma retrieval with provenance, OWASP Top 10 for LLM Applications.
Additional costs: Cloud and tool usage in your own accounts may incur charges. Set a budget and clean up after testing.
Work in your own accounts and repositories. Never share credentials or confidential data.
What’s included
Project brief, target-design guidance, assigned tasks, resource guidance and interview prompts. Detailed requirements become part of your purchased workspace.
P
MEET YOUR PROJECT EXPERT
Prepmeets
Project studio, PrepmeetsWe are the team that builds Prepmeets. Between us we have run cloud platforms, delivery pipelines, analytics teams and AI governance programmes in banks, hospitals, retailers and government departments, and these projects are the work we used to hand to new joiners in their first fortnight. Every one of them is a real scenario with the client’s name changed, scoped to a week of evenings, and reviewed by the person who originally did the job.
View verified expert profile Expert review is an optional paid service, separate from project access.
£19.90Start this project