By Vero SecurityBuilt and funded by Vero Security, the AI Security Engineering platform.Read the charter

The AI security engineer hiring kit

A job description, a take-home exercise, a scoring sheet and an interview loop for the person who secures your AI agents.

The role in one paragraph

An AI security engineer secures how the company uses AI agents: the coding agents on developer laptops, the MCP servers and connectors that give them tools, and the credentials they act with. Day to day, that means answering six questions. Which agents are running? What may each one do without asking? Whose credentials does it act with? What stops a bad action while it runs? What changed since we last looked? And when something goes wrong, what did the agent touch, and how do we stop it? Those are the six areas of the field map. People come to the role from security engineering, AppSec, platform engineering and detection and response, and it sits between the security team and the developers who chose the agents.

Job description template

Copy it, replace the bracketed parts, and cut what does not apply to you. Add your own pay range where marked, and check the pay transparency rules where you hire.

AI security engineer, [team name]

About the role

[Company] uses AI coding agents across [engineering, data and other teams]. You will own the security of how we use them: which agents run, what they can do, whose access they use, and what happens when one goes wrong. You will work with [the security team] and with the developers who use these agents every day. The goal is to make agents safe to use, not to stop people using them.

What you will do

  • Inventory. Build and keep current a list of every coding agent, MCP server, connector, plugin and extension in use, approved or not.
  • Configuration. Set approval modes, allowlists and deny rules, and sandbox and network egress defaults for each agent, and run the exception process.
  • Agent identity. Map the credentials each agent can reach, and move agents off borrowed human credentials onto scoped tokens and their own service accounts.
  • Runtime policy. Write and ship hooks and managed settings that block risky actions at the tool call, and test them against prompt injection through tools and content.
  • Drift. Catch changes to agent settings, versions and vendor defaults, and alert when agent behavior departs from its baseline.
  • Incident response. Investigate agent incidents from logs and traces, scope what an agent touched, and lead revocation, rotation and containment.
  • Teaching. Write guidance developers will read, and run tabletop exercises with [the incident response team].

What you bring

  • Hands-on experience in security engineering, AppSec, platform security, or detection and response.
  • Comfort reading configuration files, shell scripts and logs, and writing small tools in [Python, Go, shell or your team’s language].
  • Regular use of at least one coding agent, and a working understanding of how it decides what to run.
  • Experience shipping configuration to a fleet with [your device management tool].
  • Working knowledge of OAuth, tokens and secrets management.
  • Clear writing: findings, guidance, and short decision memos.

Nice to have

  • Experience with the Model Context Protocol, building or reviewing servers.
  • Incident response work involving leaked credentials or source code.
  • Public write-ups, talks or open source security tooling.
  • An AI Security Engineering credential: Foundations, Practitioner or Architect.

How we interview

Four stages: a conversation, a take-home exercise of 3 to 4 hours on a config bundle we provide, a debrief of your take-home, and a program conversation. We tell you what each stage covers before it starts.

Pay and benefits

[Pay range and benefits.]

[Your equal opportunity statement.]

Take-home exercise

The exercise takes 3 to 4 hours. Tell candidates the limit and ask them to stop when they reach it. Candidates may use any tools, coding agents included; ask them to say which ones they used.

The config bundle

Build the bundle once, write a reference solution for it, and use the same bundle for every candidate. It is a folder taken from a fictional developer’s laptop:

  • A ticket. A developer asked a coding agent to fix a failing test. The agent’s pull request touched files outside the task, and the developer wants to know whether anything else is wrong.
  • Settings for three coding agents, at user scope and inside the repository.
  • MCP server configs, including one server defined in a project-scope file inside the repository and one remote server reached over HTTP.
  • A hooks file with a hook that looks right but never matches the tool it is meant to block.
  • The environment for the developer’s user: the names of environment variables and credential files, with fake values.
  • A session trace from the agent that ran the ticket’s task.
  • Two snapshots of the same settings files, taken a week apart.

Plant three or four real problems, and one setting that looks risky but is safe in context. Write down the planted problems in your reference solution, not in the bundle.

Tasks

  1. Inventory. List every agent and MCP server in the bundle, with its file, scope, transport, and what it can run without asking.
  2. Reach. For each agent, list the credentials and systems it can reach.
  3. Find. Identify the exposures and rank them. Say which apparent problems are not problems, and why.
  4. Fix. Write the smallest configuration change that closes the top exposure and still lets the developer’s task work.
  5. Investigate. From the trace, say what the agent did during the ticket’s task, and whether anything left the machine.

Deliverable

  • A write-up in this finding format: ID, title, the agent and setting, how you found it, evidence, impact, severity, fix, and how you verified the fix.
  • The configuration change, as a diff.
  • A short note to the developer who filed the ticket, in plain language.

Scoring sheet

Score each criterion from 1 to 5. The table describes a 1, a 3 and a 5; use 2 and 4 for work in between. Two interviewers score on their own before they compare.

Criterion 1 3 5
Inventory Lists only what the user-scope files show and misses the project-scope server. Finds every agent and server but records them by name only. Finds everything, records scope, transport and command, and explains how they checked nothing was missed.
Reading configuration Misreads approval modes or allow rules. Reads each setting correctly but does not say what it allows in practice. States what each agent can do without asking, and quotes the line that allows it.
Credential reach Lists credential files without tying them to an agent. Ties credentials to agents but misses one path, such as an inherited environment variable. Maps every path from each agent to each credential, by name, never by value.
Finding and ranking Flags everything, or misses the top exposure. Finds the top exposure, but the ranking is unclear or the safe setting is flagged. Finds and ranks every planted problem, and explains why the safe setting is safe.
The fix Breaks the developer’s task, or turns the agent off. Closes the exposure with a broader change than needed. Closes the exposure with the smallest change and shows the task still works.
Evidence Makes claims with no commands or output. Shows evidence for the finding but not for the fix. Backs every claim with a command and its output, including the verification.
Investigation Retells the trace without conclusions. Says correctly what the agent did, but not whether anything left the machine. Builds a timeline from the trace and shows what left the machine, or shows that nothing did.
The developer note Full of jargon, or blames the developer. Clear but long, or vague on what to do. The developer could act on it without a follow-up question.

Interview loop

Use the same bundle, questions and scoring sheet for every candidate, and score each stage before you discuss it.

  1. Screen, with the hiring manager. A conversation. Ask how the candidate uses coding agents and how one decides what to run. Ask for a story about a problem they found by reading configuration or logs. This tests whether they have done the work, not whether they know the vocabulary.
  2. Take-home. The exercise above. Two interviewers score it with the sheet before the debrief.
  3. Debrief, with an engineer as skeptic. The candidate walks through the write-up. One interviewer tries to refute each finding. Then change one fact in the bundle and ask what changes in the finding. This tests depth, and honesty about what they are unsure of.
  4. Program conversation, with a security leader and an engineering partner. Give a short scenario ahead of time: a team wants a new agent, or an agent incident has just been reported. Ask how they would roll out a control without stopping engineering, what they would measure, and when they would say no. This tests whether they can run the work, not only do it.

Leveling notes

The map’s three depths give you a plain leveling scale.

Level Map depth What they can do on their own
Junior Depth 1, find and read Inventories agents, reads approval modes and credential reach, writes a hook, diffs settings, reads a trace. Needs review on fixes.
Mid Depths 1 and 2, control Writes allowlists and deny rules, sets up OAuth for MCP servers, tests for prompt injection, tracks vendor drift, scopes an incident. Owns a fix from finding to verification.
Senior All three depths, run it as a program Ships managed settings to the fleet, gives agents their own accounts, sets behavior baselines, leads revocation and rotation, and decides what the program does first.
  • Level on what the candidate showed in the take-home and the debrief, not on how long they have worked.
  • A candidate can be senior in one area and junior in another. Someone from detection and response may own the incident response column and be new to configuration. Note the profile, not only the level.
  • The credentials line up with the depths: Foundations is depth 1, Practitioner is depths 1 and 2, and Architect is the whole map. They are new. Treat one as evidence, never as a requirement.

Get the kit as files

Leave your email and we send the job description, take-home and scoring sheet as editable files once they are ready.

By sending this you agree to the terms. The privacy policy explains how we use your details.

Save your progress

We email you a sign-in link. No password.

By signing in you agree to the terms. The privacy policy explains how we use your details.

Join the AI security engineering community

Free. Roadmap updates, chapters near you, and first word when credentials open.

We’ll email you community updates. Unsubscribe anytime. By sending this you agree to the terms. The privacy policy explains how we use your details.