Ridgeline Certification

RCAE · AI security engineering

Ridgeline Certified AI Security Engineer

Demonstrate your ability to secure applications and agents built on large language models: model their threats, defend against prompt injection and unsafe output, contain what agents and their tools can do, protect retrieved and remembered data, manage the AI supply chain, test systems before release, and respond when they are misused.

Questions
60
Time
150 min
Try the sample case study

What RCAE validates

An AI system is secure when what it can do stays inside what it should do, whatever text reaches the model. That takes finding every route untrusted content takes into the context, putting authorization where the model cannot argue with it, limiting what agents and their tools can reach, and reading evaluations for what they actually show. RCAE assesses that judgment across the AI systems of one enterprise.

Designs and their evidence

You read architectures, tool definitions, agent and tool-call logs, retrieval settings, model files, evaluation results and red team reports, and decide what they show.

Current attacks

Indirect prompt injection, unsafe output handling, confused-deputy agents, poisoned MCP tool descriptions, multi-agent injection, poisoned knowledge bases, model files that run code, invented packages, extraction and poisoning.

Controls outside the model

Answers turn on limits the model cannot change: authorization in code, scoped identities, schemas, sandboxes and approvals.

Honest about evidence

Some questions are answered by saying what a log, a sample or a vendor's statement cannot establish.

What a holder has shown

  • The holder finds every route untrusted content takes into a model, including retrieved documents and tool results.
  • They put authorization and limits in code and identity, not in prompts.
  • They scope agents, tools and MCP servers to their task, and contain what they can do when steered.
  • They read evaluations and red team results for what they prove, and turn findings into lasting tests.
  • They passed a timed, open-book, unsupervised online exam under the candidate agreement within the last three years.

Who it is for

AI security engineers, application security engineers and developers who build or review LLM applications and agents. The exam is fair to a practitioner from any stack: models, frameworks and providers vary from case to case, and where a product's behavior matters, the question supplies it.

There are no prerequisites and no required course: anyone with a voucher can sit the exam. It is written for people who already build or review LLM applications and agents.

The exam experience

Most of the exam is case studies, all set in one multinational group's AI systems. Each gives you what an engineer would have at that point: the design, its logs and results, and the business's constraints.

Inside a case study

  1. The brief. Tabs for the Organization, the Environment, What has happened so far, and the Requirements and constraints. The behavior of the systems a question turns on is stated there, and the constraints often decide between answers.
  2. The exhibits. Four: architectures, tool definitions, agent and tool-call logs, retrieval settings, model and package records, evaluation results and red team reports. Values agree across a case's exhibits.
  3. Questions following the work as it unfolds, with the brief and exhibits beside each one.

To see how one reads, work through the sample case study, with its answers explained.

What you work from

ActivityAreaWhat you work from
A threat model before a new toolAI system threat modelingA design, content sources, a workshop list and a beta log
A coding assistant's agent modeAI system threat modelingA design, a developer's laptop, pilot sources and notes
Instructions in an order notePrompt injection and unsafe outputTool calls, the note, controls in place and proposed fixes
Model output feeding other systemsPrompt injection and unsafe outputPenetration test findings, output paths, a wiki page and proposals
A service desk agent anyone can askAgents, tools and MCP serversAn agent design, actions, credentials and tickets
MCP servers on developer laptopsAgents, tools and MCP serversAn inventory, a tool description, a network scan and logging
Two SOC agentsAgents, tools and MCP serversAgent designs, actions, an email subject and notes
An assistant that reads everythingData in retrieval and memoryA design, a red team test, sharing and logs
Memory and a shared cacheData in retrieval and memoryMemory design, complaints, the model provider and logging
A model from a public hubAI supply chain and model riskA request, a file scan, a suggested package and build logs
A model offered to partnersAI supply chain and model riskAn API, request volumes, retraining and a partner
A release gate that passedEvaluation and red teamingEvaluation results, the gate, the suites and controls
An automated red team reportEvaluation and red teamingA summary, the method, a manual review and a draft line
An agent used against an executiveMonitoring, response and governanceA timeline, a ticket, the agent and what is recorded
An AI governance programMonitoring, response and governanceAn inventory, a meeting assistant's terms, a claim and a plan

Evidence

  • Architectures, prompts and tool definitions
  • Agent actions and tool-call logs
  • Retrieval sources, indexes and memory
  • Model files, package records and build logs
  • Evaluation results and red team reports
  • Incident timelines, vendor terms and inventories

Any model, any stack

Models, frameworks and providers vary from case to case. You read designs, logs and results; you never write code or prompts. Where a product's behavior matters, the question states it.

Domains and weighting

The exam is scored in seven areas, and your result reports your score in each. The skills measured list what each area covers.

AI system threat modeling

10–15% of the exam

Maps an AI system's components and data flows, finds where untrusted content enters, and ranks threats by what an attacker gains.

Skills measured
  • Map an AI system's architecture
  • Threat model AI systems

Prompt injection and unsafe output

15–20% of the exam

Recognizes direct and indirect prompt injection, judges which controls reduce it, and makes model output safe for what consumes it.

Skills measured
  • Defend against prompt injection
  • Handle model output safely

Agents, tools and MCP servers

20–25% of the exam

Chooses how agents authenticate and what they may do, secures tools and MCP servers, and contains agent actions.

Skills measured
  • Design agent identity and authorization
  • Secure tools and MCP servers
  • Contain agent actions

Data in retrieval and memory

10–15% of the exam

Keeps retrieval within each user's permissions, governs what is indexed, and isolates memory, caches and customers.

Skills measured
  • Secure retrieval
  • Protect sensitive data and tenants

AI supply chain and model risk

10–15% of the exam

Verifies models, packages and AI-written code before they run, and defends models against extraction and poisoning.

Skills measured
  • Secure the AI supply chain
  • Manage AI-written code
  • Defend models against adversarial ML

Evaluation and red teaming

10–15% of the exam

Designs evaluations and release gates, and reads evaluation and red team results for what they prove.

Skills measured
  • Evaluate security before release
  • Red team AI systems

Monitoring, response and governance

10–15% of the exam

Monitors AI systems, contains and investigates AI incidents, and governs AI use across the organization.

Skills measured
  • Monitor AI systems
  • Respond to AI incidents
  • Govern AI security

Exam details and requirements

ExamRCAE-01
Questions60
Duration2 hours 30 minutes
Passing score70%
DeliveryUnsupervised online, on this site, whenever you are ready. There is no proctor and no ID check; you sit it under the candidate agreement.
MaterialsOpen book: your own notes and reference documentation. No help from another person or an AI model.
RetakesTwo attempts per voucher; the next opens 24 hours after an attempt that does not pass
Validity3 years

Delivery and integrity

  • Before the exam opens you accept the candidate agreement: identity, independent work, permitted resources, confidentiality and integrity.
  • Each attempt has its own set of questions, in its own order.
  • The clock runs on our server. A dropped connection loses no answers, but the clock keeps running.
  • Answers are saved as you give them. You can flag questions and return to them from a review screen.
  • You can report a problem with any question from the exam screen, and every report is reviewed. Results can be appealed under the appeals procedure.

Prepare and maintain

The exam study guide covers how to practice for each kind of question and what to expect on the day. Courses are optional; these cover the exam's skills.

Core courses

Recommended

Examined, but not required.

Voucher

$349

Two attempts at the exam. A bought voucher does not expire.

Payment by card through Stripe. The voucher is added to your account as soon as the payment is confirmed. Refunds follow the Refund Policy.

Included with a yearly training plan

A yearly Ridgeline training plan, and each Business seat, includes one exam voucher a year. Link your training account to this site to claim it. It is usable while the plan renews and expires when it ends. Monthly subscribers buy a voucher like anyone else.

Renewal

RCAE is valid for 3 years from the day it is issued. Renew it by passing the current version of the exam in the last twelve months before it expires; the renewed certification runs 3 years from your current expiry date, so renewing early loses no time. A renewal voucher of two attempts is free with a yearly training plan or Business seat and $49 otherwise. We email you 90 and 30 days before expiry. An expired certification still verifies, shown as expired, and holding RCAE again then takes a full voucher.

Verification

Every RCAE has a public verification page showing the holder, the dates and whether it is valid, expired or revoked. The rules behind the certification are in the published policies.

Questions

Is it tied to one model provider?

No. Cases use different models and frameworks, and where a product's behavior matters, the question states it.

Do I need to write prompts or code?

No. The exam shows designs, logs and results and asks what they show and what you would do.

Is there a hands-on lab?

No. The evidence a system would produce is in the exhibits.

Can I have extra time or another adjustment?

Adjustments, including extra time, are agreed case by case under the reasonable adjustments policy. Ask before you start an attempt.

What do I receive when I pass?

A certification ID, issued at once in the name on your account, with a public verification page that shows its status and dates.

Can I use documentation during the exam?

Yes, your own notes and documentation. You may not get help from another person or an AI model.

Ready to sit RCAE?

Read the study guide, then sit the exam when you are ready.

Exam study guide