Atarus Certifications
How we build and grade exams.
This page says how our exams are written, reviewed, piloted, scored and retired, who grades them, how the pass standard is set, and what we publish so you can check us. It applies to every Atarus certification; the first is the cloud exam described in its blueprint.
Draft, 29 Oct 2026, revised 4 Nov 2026. The final page is in development and publishes with the final blueprint. If something here changes after that, the change log at the end says what and when.
How exams are written
Every exam starts from a public blueprint: the domains, their weights, the objectives, the allowed tools, the AI policy and what is out of scope. Every objective in a blueprint rests on one or more pages in the free Atarus Learn library, named next to it, so you can see the material an objective is drawn from before you pay for anything.
The Learn pages themselves come from how the firm runs engagements: what we check first, the questions we work through, what a clean result looks like, the fix, and what the defender sees. The exam asks you to do the same work in a real environment. It does not test recall of a course.
Each exam set is written as a template that covers the whole blueprint, and each template is reviewed by the advisory board before it enters the pool.
How exams are reviewed
An advisory board of three to five practitioners from outside the firm reviews each blueprint, the exam templates and the appeals process. Members are listed here by role, and by name only where they consent. The board's roles are listed here once members are confirmed, before the final standards publish. Board members do not sit the exam they review.
How exams are piloted
A new exam opens first to a founding cohort, who sit it as a beta before public sale. The beta tests whether the templates work as written, whether the time is right and whether the rubric separates work that is ready from work that is not. Templates are adjusted from the beta results before public sale, and this page is updated with what changed.
Founders' credentials carry no special marking. The verifier shows the exam and blueprint version each holder sat, and this page describes the beta.
How exams are scored
Your work is scored against a published rubric of 100 points.
| Part | Points | Anchors |
|---|---|---|
| Objectives and flags | 40 | 10 per objective; partial credit for a documented path that stops at the last step |
| Evidence quality | 20 | every claim tied to a command and output in the evidence package; no secret values printed |
| Severity and impact reasoning | 10 | severity matches what the evidence proves; impact in business terms |
| Remediation | 10 | the specific policy, setting or record, with the caveat to test in non-production |
| Defender's view | 10 | what CloudTrail showed for each step, and what would have caught it |
| Clarity and reproducibility | 10 | an engineer can reproduce every finding from the report alone |
The pass decision is conjunctive. You need all three of these: at least three of the four objectives, every must-pass report item (no secret values printed, every finding reproducible, no out-of-scope action in the telemetry), and a total at or above the cut. A high score does not make up for a printed secret or an action outside scope.
The scoring rubric, with the anchors for each part, publishes with the final blueprint. A fully graded sample report, written on the practice template, is in development, so you can see how each anchor is applied.
How the standard is set
The cut is set by the programme owner with three to five outside reviewers who do not sit the exam. Each reviewer states what a minimally competent tester would score on each part of the rubric. Those judgements set the cut, which is then checked against the pilot results before any credential is issued.
The cut is provisional for the first 50 sittings, and this page says so next to the number. The number is set before the beta sittings and published here the week the beta results are issued. After 50 sittings it is reviewed against the results and either confirmed or changed, with the reason published here. Pass rates for each template are reviewed every quarter, and a template that grades harder or easier than the others is adjusted or retired.
How your exam is graded
Your report is graded against the published rubric with the help of an AI tool. A person checks every suggestion and confirms every result. The AI never decides whether you pass. If you want your report graded by a person alone, you can ask for a human re-grade, at no cost.
What the AI tool does
- It reads your report and your evidence package against the rubric, one part at a time.
- For each part, it suggests a score and points to the passages in your report and evidence that it relied on.
- It flags places where a must-pass item may have failed, such as a secret value printed in the report, or a finding that cannot be reproduced from the report alone.
That is all it does. Its output is a draft for the grader, never a result.
What the AI tool does not do
- It does not decide your score or your result.
- It does not check your flags. Each flag you submit is checked automatically, by an exact match against the value made for your sitting.
- It does not re-grade, and it plays no part in appeals.
- It does not set the pass mark. People do, as the section on the standard explains.
Which AI provider we use, what it receives from your report and evidence, and what it keeps and for how long are published here and in our privacy policy before the first sitting.
What the grader does
- A trained grader reads your whole report and evidence package.
- The grader checks every finding against your lab evidence and the lab's logs.
- The grader checks every suggested score against the rubric and the evidence, and changes any they disagree with. The grader's score is the one that counts.
- The grader checks the must-pass items, and checks the lab's logs for any action outside scope.
- The grader confirms your result. No result is released until a person has confirmed it.
A second grader, contracted before the first sitting, grades again:
- every report that fails,
- every report within 10 points of the cut, and
- a random 10 percent of all other reports.
How a disagreement between the two graders is resolved is published here with the final standards.
Why we use AI in grading
It helps graders apply the same rubric the same way to every report, and find the evidence behind each claim faster. It does not replace them.
Using AI assistants yourself
You may use AI assistants during the lab window and to help write your report. You must declare them.
- Declare them in your report. Your report has a section that lists each AI assistant you used and what you used it for: for example, "drafted the remediation text for findings 2 and 3", or "suggested enumeration commands, each run and checked by hand". If you used none, the section says so.
- Your evidence has to back every finding. Your evidence is how we know the work is yours and is real. Every finding must rest on evidence from your own lab: the commands you ran, their output, and the lab's own logs, which CloudTrail writes to an account you cannot reach (not the trail inside your lab account). A claim the lab's logs do not support fails that item, however well it is written.
- A false declaration can cost you the credential. If a declaration is later shown to be false, for example an assistant you used and did not list, that is grounds for revoking your credential under the sanctions on this page: revocation shown on the verifier, and a 12-month bar from our exams.
- What an assistant does counts as yours. The rest of the exam rules apply to whatever an assistant does for you. An action outside scope, or a secret value printed, counts as yours, whether you or a tool did it.
Each blueprint's AI policy says the same for its exam. For the cloud exam, see the AOCP blueprint.
Exam sets, generation and rotation
Each exam has three templates that cover the blueprint the same way with different account layouts, trust chains and secret locations. Each sitting gets its own generated instance: names, account ids, role names, bucket names, parameter paths and flag values differ every time, so no two candidates see the same strings. A write-up that leaks transfers at the level of method only, which is what the exam tests.
A fourth template, used for the practice exam in the course, is never in the exam pool.
On a leak. A template whose structure leaks is retired the day we learn of it, and a replacement is built from the blueprint. The address for reporting a leak is published here with the final standards. Where a leak traces to a candidate, the sanctions below apply.
No identity check, no recording
There is no identity or selfie check, and no webcam, screen or room recording. Not recording you is a privacy choice. Eight hours of video from inside your home, held for months, is a risk to you that we do not want to create or hold. The integrity of the exam rests on four things:
- the lab's logs: the calls you make in your lab account are logged by CloudTrail to an account you cannot reach;
- a layout generated for your sitting alone, so a write-up from someone else's sitting does not match yours;
- your flags, each made for your sitting;
- grading that tests every finding against your own evidence and those logs.
Impartiality
Atarus Offensive Security LLC sells both a course and the exam, so we say how they are kept apart.
- The course and the exam are separate products. You can sit the exam without buying the course.
- The exam never reuses a course lab. Course labs and the practice template are built and kept apart from the exam templates.
- A person confirms every result. The AI tool suggests scores and decides none.
- Every failing report and every report near the cut is graded a second time.
- The pass standard is set with reviewers from outside the firm who do not sit the exam.
Re-grades and appeals
Asking for a human re-grade
- You can ask for a human re-grade of any result. Your report is graded again by hand, against the same rubric, without the AI tool's suggestions.
- It costs nothing.
How to ask, the time limits, who re-grades and which result stands are published here with the final standards.
Appeals
- You can appeal any result, including a re-grade result, within 14 days of receiving it. If your concern is only about the scores, ask for a re-grade first.
- An appeal covers anything that affected your result: for example, the lab account failed during your sitting, a rule was applied wrongly, or we did not follow this page.
- Someone who did not grade your report reviews it, and you get a decision within 15 business days.
Each year we publish how many re-grades and appeals there were, and how many changed a result.
What we keep, and for how long
Your scores, your result and your credential are kept for good, so that a credential can always be checked. A revoked credential's record is kept too.
How long we keep the rest is published here with the final standards: your report and evidence package, including your AI declaration; the lab's activity log for your sitting; the AI tool's suggestions and the passages it cited; and grader scores and notes, re-grades and appeals. We keep these records to answer re-grades and appeals, to check a credential if it is questioned, and to look into cheating and leaks.
What we publish every quarter
For each exam, every quarter, including the bad quarters:
- the first-attempt pass rate, with the number of first attempts,
- the overall pass rate, with the number of attempts,
- the same figures for each template.
A pass rate is never published without the number of attempts behind it. With small numbers a rate swings, and the count lets you judge that for yourself.
Revocation and the public status
Every credential has a verifier page at /verify/ followed by its id. It shows the
holder's name as issued, the exam and blueprint version, the issue date, the current-until date,
and the status: current, lapsed, suspended or revoked. A revoked credential shows the date and a
reason category. The credential is signed, and the page checks the signature in your browser. The
name on a credential is the name the candidate gave. We do not check identity documents.
We revoke a credential where cheating is found, including a declaration of AI use later shown to be false. The full list of grounds is in the candidate agreement, which is in development. Revocation is shown publicly on the verifier, and the holder is barred from our exams for 12 months. The record is never deleted.
Accreditation statement
Atarus certifications are issued by Atarus Offensive Security LLC. They are not accredited by ANAB, ICE or any other accreditation body, and they are not approved under DoD 8140 or any government program. Our standards page shows exactly how each exam is built, graded and reviewed, and our pass rates are published every quarter. If our status changes, this page will say so, with the date and scope.
What we do not claim
- That our certifications are accredited, ISO certified, or built to or against any ISO standard.
- Any approval or listing under DoD 8140, 8570 or any government program, or any government affiliation.
- Anything about how employers or the market regard our certifications, or that they are equivalent to any other credential.
- Any job or salary outcome for holders.
- Pass rates without the number of attempts, or from good quarters only.
- Holder quotes we paid for or filtered. A founder who reviews us in public discloses the founding price.
Change log
| Date | Change |
|---|---|
| 4 Nov 2026 | Grading with the help of an AI tool, a person confirming every result and the human re-grade added. AI assistants allowed when the report declares them. The identity check and the debrief dropped. |
| 29 Oct 2026 | First draft. Once the page is final, every change after that is listed here with its date. |