CASE STUDY · AI AGENT SECURITY & GOVERNANCE
Auditing an AI agent fleet before a client's SOC 2 review.
- 1818 high-priority findings closed before the audit window
01 · PROJECT OVERVIEW
The client and the brief.
An anonymised mid-market B2B services company. Name withheld under confidentiality.
An agent inventory, permission review, added logging and a named owner for each agent.
- 13 agents in use, no inventory, and a SOC 2 audit in 8 weeks.
- 2The agents had been set up by different teams for different jobs.
- 3One drafted replies to support tickets.
- 4One summarised sales calls.
- 5One moved data between two internal tools.
02 · WHAT WENT WRONG
What went wrong, and how we fixed it.
Every engagement hits problems. These were the ones that mattered here, and what we changed so they did not come back.
- PROBLEM 01Midway through, we found a fourth agent that nobody had mentioned
A team had built it with a no-code automation tool and a personal account. The inventory had missed it because it did not live in any repository. We added a platform-level check of the automation tool's connected apps and a short form that any team must complete before connecting an agent to company data.
- PROBLEM 02The unlisted agent was brought under an owner and a service account
- PROBLEM 03A second hurdle was ownership
Two agents were built by people who had since moved teams, so the natural owner had left. We asked each department head to name a current owner, wrote the duty into their objectives and recorded a backup. One agent had no clear business use any more and was switched off, which removed a set of credentials from circulation.
- PROBLEM 04One agent had no clear business use any more and was switched off, which removed a set of credentials from circulation
03 · THE PROCESS
How we ran it.
We split the 8 weeks into four blocks.
- 01Weeks 1 to 2
inventory. We interviewed the teams, read the configuration of each agent and listed every credential, data source and action it could use.
- 02Weeks 3 to 4
permission review and threat testing. We compared what each agent could do with what it needed to do, and tried to push each one outside its job.
- 03Weeks 5 to 6
remediation. We removed excess access, added logging and approval steps and assigned owners.
- 04Weeks 7 to 8
evidence and rehearsal. We collected screenshots, logs and policies in one pack and walked the security lead through the questions an auditor would ask.
We met the security lead weekly and posted a running findings list that she could read at any time.
04 · RESULTS
What changed.
Measured against the way the work ran before the engagement began.
- NOW 01
No agent goes live without an entry in the register and a named owner.
- NOW 02
Permissions are reviewed on a fixed schedule.
- NOW 03
The security lead can answer "what can this agent do and who is responsible?" from one document.
- NOW 04
New agent requests go through the same short form, so the register stays current.
- 18
18 high-priority findings closed before the audit window
No critical findings open at the end of the pilot
05 · WHAT WE BUILT
How it works.
- The inventory records, for every agent, the owner, purpose, model, tools, credentials, data accessed and systems it can change.
- Permissions were reviewed against the task.Shared API keys were replaced with scoped service accounts so that each agent has its own identity in the logs.
- Logging captures the input, the tool call, the result and the time, and is written to storage the agents cannot edit.
- Actions with side effects outside the company, such as sending messages, now need a person to approve.
- Findings were ranked by severity and each one was tracked to closure.
- Handover
- We left a short review routine: monthly permission checks, a quarterly re-test of the riskiest agent and a change log for each agent. The security lead ran the first monthly review herself while we observed.
06 · QUICK ANSWERS
The short version.
- What problem did the client have?
- 3 agents in use, no inventory, and a SOC 2 audit in 8 weeks.
- What did Brandsmashers build?
- An agent inventory, permission review, added logging and a named owner for each agent.
- How was the work run?
- We split the 8 weeks into four blocks.
- What were the results?
- 18 high-priority findings closed before the audit window; no critical findings open at the end of the pilot.
YOUR TURN
Have a similar requirement?
Tell us what is stuck. We will come back with a plan and matched, pre-vetted engineers.
