Ridgeline Certification

RCAE · Official exam study guide · RCAE-01

Ridgeline Certified AI Security Engineer

What the RCAE exam assesses, how it is delivered, and how to prepare for it, with or without Ridgeline training.

Introduction

The Ridgeline Certified AI Security Engineer (RCAE) exam assesses your ability to secure applications and agents built on large language models: modeling their threats, defending against prompt injection and unsafe output, containing agents and their tools, protecting retrieved and remembered data, managing the AI supply chain, testing before release, and responding to misuse.

Each question gives you a design, a log or a result and asks what it shows or what you would do. The exam has no hands-on lab: it tests the judgment you bring to AI systems.

Who should sit it. 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.

Recommended experience. Working experience of building, reviewing or securing LLM applications or agents. No course is required to sit the exam.

Exam specifications

ExamRCAE-01
Questions60
Duration2 hours 30 minutes (150 minutes), on one clock for the whole exam.
Passing score70%: 42 of 60 correct. Every question carries equal weight.
DeliveryUnsupervised, online, on this site, in a desktop browser. There is no proctor and no identity document check; see exam security.
MaterialsOpen book: your own notes and reference documentation. No help from another person or an AI model.
AttemptsTwo per voucher; 24 hours between an attempt that does not pass and the next
Validity3 years from the day the certification is issued

Assumed knowledge

  • 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. Where a product's behavior matters, the question or its exhibits supply it.
  • Frameworks. OWASP Top 10 for LLM Applications, MITRE ATLAS, NIST AI RMF and ISO/IEC 42001 at the level of their ideas.
  • Platforms. No question depends on one model, provider or framework.

Skills measured

The same list is published on its own page to print or save: RCAE skills measured.

Skills measured as of November 21, 2026.

Audience profile

You secure applications and agents built on large language models. You work out where untrusted input reaches a model, contain what agents and their tools can do, protect the data a model retrieves and remembers, check the models, packages and AI-written code a system depends on, test systems before release, and detect and respond when an AI system is misused.

You work with application developers, platform and identity teams, the SOC, and the people who govern AI risk. 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.

Skills at a glance

  • Model AI system threats (10–15%)
  • Defend against prompt injection and unsafe output (15–20%)
  • Secure agents, tools and MCP servers (20–25%)
  • Protect data in retrieval and memory (10–15%)
  • Manage AI supply chain and model risk (10–15%)
  • Evaluate and red team AI systems (10–15%)
  • Monitor, respond and govern (10–15%)

Model AI system threats (10–15%)

Map an AI system's architecture

  • Identify the components of an LLM application, including model, orchestration, retrieval, tools, memory and user interface
  • Trace data flows between users, the model, retrieved content and tools
  • Identify where untrusted content enters the context window

Threat model AI systems

  • Identify trust boundaries in an AI system
  • Map threats to OWASP Top 10 for LLM Applications and MITRE ATLAS
  • Prioritize threats by what an attacker gains at each boundary

Defend against prompt injection and unsafe output (15–20%)

Defend against prompt injection

  • Distinguish direct from indirect prompt injection
  • Identify injection carried by documents, web pages, email and tool results
  • Evaluate controls against injection, including instruction hierarchy, input separation and output checks
  • Identify which controls reduce injection and which only appear to

Handle model output safely

  • Treat model output as untrusted input to downstream systems
  • Prevent model output reaching HTML, SQL, shell or code execution unchecked
  • Validate structured output before it drives an action

Secure agents, tools and MCP servers (20–25%)

Design agent identity and authorization

  • Select between an agent acting with a user's delegated access and acting as its own identity
  • Scope an agent's permissions to its task
  • Manage non-human identities and their credentials

Secure tools and MCP servers

  • Evaluate a tool's or MCP server's permissions, inputs and trust
  • Restrict which tools an agent may call, and with what arguments
  • Identify tool poisoning and malicious tool descriptions
  • Secure MCP server authentication and transport

Contain agent actions

  • Require human approval for high-impact actions
  • Limit an agent's reach, including rate, scope and spend
  • Isolate code an agent runs
  • Contain multi-agent systems, where one agent's output becomes another's instructions

Protect data in retrieval and memory (10–15%)

Secure retrieval

  • Control what content is ingested for retrieval and who can change it
  • Enforce the user's permissions on retrieved content
  • Identify poisoned or over-shared content in a knowledge base

Protect sensitive data and tenants

  • Prevent sensitive data reaching a model, logs or a provider it should not
  • Isolate tenants in shared AI applications
  • Secure agent memory against poisoning and leakage
  • Apply privacy-preserving techniques where data must be used

Manage AI supply chain and model risk (10–15%)

Secure the AI supply chain

  • Verify the provenance of models, datasets and packages
  • Identify risks in model formats and model loading
  • Evaluate a third-party model or AI service before adoption

Manage AI-written code

  • Review AI-written code for security defects
  • Identify packages an AI assistant invented or misnamed

Defend models against adversarial ML

  • Identify evasion, poisoning, extraction and inversion attacks against the company's own models
  • Select defenses in proportion to a model's exposure

Evaluate and red team AI systems (10–15%)

Evaluate security before release

  • Design security evaluations for an AI system
  • Set release gates from evaluation results
  • Read evaluation results honestly, including pass rates and their limits

Red team AI systems

  • Plan automated and manual red teaming
  • Interpret red team findings and decide what they prove
  • Turn findings into fixes and regression tests

Monitor, respond and govern (10–15%)

Monitor AI systems

  • Select telemetry for AI systems, including prompts, tool calls and decisions
  • Detect misuse, injection attempts and abnormal agent behavior

Respond to AI incidents

  • Contain an AI incident, including revoking an agent's access and disabling tools
  • Investigate an AI incident from its logs, and state what they cannot show

Govern AI security

  • Map AI security controls to NIST AI RMF and ISO/IEC 42001
  • Maintain an inventory of AI systems and their risks
  • Assess third-party AI services and agents

The skills listed under each objective illustrate how it is assessed; related topics may be covered. Where a product's behavior matters to a question, the question or its exhibits state it.

Change log

BeforeNowChange
All skill groupsFirst published

Question formats

Individual questions

The first 12 questions stand alone. Most carry no exhibit and ask for a judgment AI security engineers make often; some carry one: a log, a result or a set of values.

Case studies

Each case study is one situation that unfolds over several questions: instructions in an order note, model output feeding a console, an agent anyone can ask, a poisoned tool description, an assistant that reads everything, a model file that runs code, a release gate that passed. It opens with a brief in tabs (Organization, Environment, What has happened so far, Requirements and constraints), then four exhibits. The brief and exhibits stay on screen beside every one of the case's questions.

The brief matters. What a feature must keep doing, and what an agent may never do, often decides between two answers that are otherwise sound.

The sample case study is a complete case in this format, with the answers and the reasoning behind each option. It is not in the exam.

Preparing

The table suggests one way to practice each area and what readiness looks like. It is a guide, not a syllabus: the skills measured above are what the exam assesses.

AreaPracticeYou are ready when
AI system threat modelingDraw the data flows of an AI application you know, marking every source an outsider can write to.You can name each route untrusted content takes into the model and rank the threats by what an attacker gains.
Prompt injection and unsafe outputPlant an instruction in a document or tool result and see whether an assistant follows it.You can say which controls reduce injection and which only appear to, and make output safe for its consumer.
Agents, tools and MCP serversList an agent's tools, the identity each acts with and what each can reach.You can scope each tool to its task and name the actions that need a person.
Data in retrieval and memoryAsk an internal assistant for content you should not see, and trace why it answered or refused.You can show retrieval stays within each user's permissions and that memory and caches are isolated.
AI supply chain and model riskInspect a model file and a set of suggested packages before using them.You can tell a weights-only format from one that runs code, and stop invented packages reaching a build.
Evaluation and red teamingRun a small injection suite against an assistant and work out the uncertainty in its pass rate.Your gate sets a bar per category and states what the evidence cannot show.
Monitoring, response and governanceWrite the logging an agent needs for an investigation, and the steps to contain it.You can cut an agent's power within minutes and say what its records cover.

Readiness check

Before you book an attempt, check that you can:

  • Find every route untrusted content takes into a model
  • Put authorization for tool actions in code, not in prompts
  • Make model output safe for HTML, SQL and actions
  • Scope agents, tools and MCP servers to their task
  • Keep retrieval and memory within each user's rights
  • Verify models and packages before they run
  • Read evaluation and red team results for what they prove
  • Contain an AI incident and say what the logs cannot show

Confidence is not a substitute for practice: work through the sample case study without looking at the answers first.

Limits the model cannot change

Instructions to a model reduce risk; they do not bound it. Several answers turn on putting the limit in code, identity or a person instead.

Reading results honestly

Pass rates, judge labels and vendor statements say less than they seem to. Say what they cannot show.

Resources

Official references

Free documentation for each part of the work. The exam assumes nothing beyond what these cover.

SubjectReference
Risks in LLM applicationsOWASP Top 10 for LLM Applications
Adversary techniques against AIMITRE ATLAS
AI risk managementNIST AI Risk Management Framework

Ridgeline courses (optional)

Ridgeline training covers the exam's skills; it is one way to prepare, never a requirement. Core: SEC502 Securing LLM Applications and AI Agents, SEC410 AI Security Engineering, plus . Recommended: GRC302 AI Governance and Risk Management. A yearly Ridgeline training plan, or a Business seat, includes one exam voucher a year.

On exam day

Time

A suggested plan, not enforced by the exam: about 24 minutes for the 12 individual questions, 116 minutes for the case studies (about 2.4 minutes a question), and 10 minutes to review flagged questions. Some cases take longer than others.

Navigation

  • Flag a question you are unsure of and come back to it from the review screen, which can step through only the flagged or unanswered questions.
  • Unanswered questions count as incorrect, so answer every one.
  • Every answer can be changed until you finish.

If your connection drops

The clock runs on our server and every answer is saved as you give it. Reload the page or sign in again and you return to the same question with your answers intact; the clock keeps running while you are away. If the time runs out before you return, the attempt ends with the answers you gave.

If something fails on our side

If a fault in the exam system affects your attempt, such as the exam not loading or answers not saving, report it from the exam screen or write to certification@ridgelinecyber.com the same day. It is handled under the appeals procedure; where we confirm a fault on our side, the attempt is replaced at no cost.

Permitted materials

  • Allowed: your own notes, your own files, reference documentation and web searches for documentation.
  • Not allowed: any AI model or AI assistant, including AI answers built into search engines and tools; any other person; anything that copies or sends the exam's content elsewhere.

Disputing a question

Report a question you believe is wrong with the link under it on the exam screen, during the exam; it does not stop the clock. After the exam, you can appeal your result on that ground under the appeals procedure. If an appeal about a question is upheld, every attempt it affected is reviewed.

Before you start

  • Check your name on your account: it is printed on your certification exactly as entered.
  • Use a desktop or laptop with a large screen. The exam screen has text size controls and a dark mode.
  • If you need an adjustment, ask before you start: see reasonable adjustments.

Results and policies

You see your result as soon as you finish: your overall score and your score in each area. After an attempt that does not pass, it stays on the exam page until your next attempt. The correct options are not shown, because the questions stay in use. A pass issues your certification at once, with an ID anyone can check at verification. It is valid for 3 years and renews by passing the current version of the exam in its last twelve months; the renewal runs 3 years from your current expiry date.

A voucher gives two attempts at the exam, and starting an attempt uses one. After an attempt that does not pass, the next opens 24 hours later with different questions.

The candidate agreement

You accept these terms on screen before each attempt (version 2026-11-16):

  1. Identity. You are the person named on this account, and you will sit the exam yourself.
  2. Independent work. You will work alone. You will not get help from another person, an AI model or an automated answer service while the exam is open.
  3. Permitted resources. The exam is open book: you may use your own notes and reference documentation, within the independent-work term.
  4. Confidentiality. The questions, case studies, exhibits and answers are confidential. You will not copy, photograph, record, reproduce, share or discuss them, during or after the exam.
  5. Integrity. You will follow the published exam rules and will not obtain or distribute unauthorized exam material. Suspected misconduct is investigated under the exam security policy and can lead to cancelled attempts or a revoked certification; every decision can be appealed.

The full rules are in the policies.

Document control

Exam versionRCAE-01
Skills measured as ofNovember 21, 2026
Candidate agreementVersion 2026-11-16

A change to the objectives, the question formats or the scoring comes with a new exam version and is announced here before it applies.