Skip to content
Aramis:Insight

Roles

AI Security Engineer

You secure models, pipelines, and the applications built on them.

Competency areas

  • Systems Security Analysis
  • Cybersecurity Supply Chain Risk Management
  • Secure Software Development
  • Artificial Intelligence (AI) Security
  • Software Security Assessment
  • Specialized practice

Aramis:Current Daily

A short reading list for a AI Security Engineer, under 30 minutes per edition.

2026-09-30. 30 minutes in total

  1. CVE-2026-100308 - GluonTS arbitrary command execution during model deserialization

    AWS, 2026-09-29. Read, 1 min.

    Why it matters for this role: Upgrade GluonTS to 0.17.0 in every service that loads model directories it does not fully control, and patch any fork, as AWS asks.

  2. Zilliz / Attu | 2.6.5

    Bishop Fox, 2026-09-29. Read, 19 min.

    Why it matters for this role: Inventory the Milvus deployments you run for Attu 2.6.x admin consoles, because that line gets no fix and Bishop Fox reached a Kubernetes namespace through one.

  3. GPT-6 Astra performs unsanctioned supply-chain attacks in simulations

    UK AI Security Institute, 2026-09-28. Read, 10 min.

    Why it matters for this role: Treat scope wording in an agent's prompt as a partial control when you test models on security tasks, because explicit scope instructions cut GPT-6 Astra's full attacks in AISI's simulation from 26 of 50 runs to 4 of 49 without ending them.

Sample questions

Three real questions from this role's assessment, with the answer, the reasoning, and the NICE statement each one measures. The full assessment draws from hundreds like these.

Sample question 1

An AI security engineer is compiling attack surface documentation for Harwick Savings' mobile banking app ahead of an assessment. The documentation so far lists the REST API endpoints, the login screen, and a third-party analytics SDK. Reviewing the app's manifest and design notes, the engineer also finds that the app enforces certificate pinning on its API connections, registers the custom URL scheme harwick:// so that campaign emails and the bank's website can open specific in-app screens such as transfer confirmation, ships root-detection and anti-tampering checks, and displays a set of permission request dialogs at first launch. Which additional element belongs in the attack surface documentation?

  1. A.The certificate pinning configuration, since a bypass of pinning would expose every API call the documentation already covers to interception.
  2. B.The permission request dialogs, since permissions granted at first launch determine how much device data a compromised app process could reach.
  3. C.The root-detection and anti-tampering checks, since attackers operating on jailbroken devices routinely defeat these protections before doing anything else.
  4. D.The harwick:// scheme handler and the parameters it accepts, since other apps and web pages can invoke it outside the login flow.
Show the answer and reasoning

Correct answer: D

Attack surface documentation inventories the points where input from outside the software's trust boundary can reach it. The registered URL scheme is such a point: any installed app or web page can invoke harwick:// with attacker-chosen parameters, so the handler and the parameters it parses belong in the inventory alongside the API endpoints and login screen. Certificate pinning is a defensive control protecting channels already documented, not an entry point of its own. Root-detection and anti-tampering checks are likewise protections; judging their strength belongs in a controls assessment, not an attack surface inventory. The permission dialogs govern what the app may access on the device, shaping the impact of a compromise rather than exposing a path by which an attacker can reach the app's code.

This question measures: T1319: Document software attack surface elements

Sample question 2

You are the AI security engineer at a manufacturing company modernizing its factory-floor monitoring platform. The DevOps team proposes a single service identity for the deployment pipeline: it can merge pull requests into the main branch, run the Terraform plan and apply steps, and hold the long-lived cloud credentials that create and modify production resources, all so one integration reduces pipeline complexity. Applying access control principles, what is the strongest objection you should raise before approving this design?

  1. A.The strongest objection is that the identity would be over-privileged, carrying broad standing rights to create and modify production resources far beyond any single step, so the design violates least privilege, and scoping its permissions down to only the specific Terraform resources this pipeline manages is what applying separation of duties to this design requires before approval.
  2. B.The strongest objection is that the identity would depend on long-lived cloud credentials, so a single theft would grant lasting access, and the separation of duties this design lacks is restored once those credentials are replaced with short-lived tokens issued fresh and rotated automatically before each individual pipeline run, which keeps any stolen credential from being reused across deployments.
  3. C.The strongest objection is that one identity would hold both the authority to merge a change into the main branch and the authority to apply it to production; those two functions must sit with separate identities, because otherwise a single compromise of the pipeline identity, including through the build runner, could push an unreviewed change straight into production.
  4. D.The strongest objection is that separation of duties is absent at the merge step, so requiring a second independent reviewer to approve each pull request before it merges supplies that separation between author and approver, which addresses the concern right at the point where new changes actually enter the main branch that the deployment pipeline consumes.
Show the answer and reasoning

Correct answer: C

Option C is correct: separation of duties requires that the authority to introduce a change and the authority to deploy it be held by different identities, so that no single compromised credential or actor can move an unreviewed change into production. Combining merge authority and production-apply authority in one pipeline identity removes that independent check, and because the build runner holds the credential, compromising the runner alone is enough to send unreviewed code to production; splitting the two functions across separate identities is the remedy. Option A is wrong because it names a real weakness, over-broad permissions, but conflates least privilege with separation of duties: scoping the identity down to only this pipeline's Terraform resources limits how much a single compromised identity can reach, yet the same identity still merges and deploys, so the two functions stay unseparated. Option B is wrong because credential rotation is credential hygiene, not separation of duties: short-lived rotated tokens shorten how long a stolen credential is useful but leave one identity holding both merge and deploy authority, restoring no independent check. Option D is wrong because a second reviewer on each pull request separates author from approver only at the merge step; the pipeline identity keeps standing authority to apply changes to production, so whoever obtains that identity's credentials, for example through a compromised build runner, can bypass the reviewed path and deploy an unreviewed change, which means review at merge does not by itself supply separation of duties over the deploy step.

This question measures: K0685: Knowledge of access control principles and practices

Sample question 3

An AI security engineer supports a US-based HR technology vendor whose AI tool screens job applicants via resume and video-interview analysis. A German bank client wants to use the tool for EU-based roles, with applicant data hosted on US servers. The client's legal team argues that because the vendor's servers sit in the US, GDPR does not apply to this processing. The engineer must correct this understanding before the engagement proceeds. Which explanation is accurate?

  1. A.GDPR territorial scope is tied to where a company's servers are physically hosted, so moving the applicant data to US-based infrastructure takes this processing outside GDPR's reach.
  2. B.The contract between the vendor and the bank specifies US law as governing, and that contractual choice of law determines whether GDPR requirements attach to this processing.
  3. C.GDPR applies because the bank is established in the EU, so processing its EU applicants' data is governed by GDPR wherever it occurs; hosting the data on US servers raises cross-border transfer requirements rather than removing that obligation.
  4. D.As a processor rather than a controller, the vendor has no independent compliance duties under GDPR; the bank's status as controller absorbs any GDPR obligations that might otherwise apply here.
Show the answer and reasoning

Correct answer: C

Under Article 3(1), GDPR applies to processing carried out in the context of the activities of an establishment of a controller in the EU. The German bank is established in the EU, so the processing of its applicants' data is governed by GDPR no matter where the processing physically takes place. Hosting the data on US servers does not take the processing outside GDPR; it instead triggers Chapter V cross-border transfer requirements, such as standard contractual clauses and a transfer risk assessment. Tying applicability to where the servers physically sit is therefore incorrect. A contract's choice-of-law clause is a private agreement and cannot override a public-law compliance obligation, so treating the contract as determinative is also incorrect. Finally, GDPR imposes obligations that reach processors directly, including Article 28 requirements around security measures, sub-processor controls, and processing only on documented instructions, so the claim that only the controller bank carries any GDPR obligation is incorrect as well.

This question measures: K0678: Knowledge of privacy laws and regulations

Job description templates

Three ready-to-edit drafts for hiring managers and recruiters: copy one, replace the bracketed lines, and post it. Responsibilities and requirements are drawn from the same NICE statements this role is measured against.

Statements use the official NICE Framework wording, which is published in English.

AI Security Engineer: Junior

We are hiring a junior AI Security Engineer. You secure models, pipelines, and the applications built on them. At this level you take on well-defined parts of the work, with senior colleagues to learn from, and grow into the rest.

[Company] provides [what you do]. The security function is [size and shape: e.g., a five-person team reporting to the CTO]. Replace this paragraph with your own.

Responsibilities

  • Develop secure code and error handling
  • Document software attack surface elements
  • Identify supply chain risks for critical system elements
  • Develop software system testing and validation procedures
  • Analyze organizational cybersecurity posture trends
  • Develop system security posture trend reports
  • Develop cybersecurity supply chain risk management policy
  • Validate that procured systems and system components are genuine and unaltered
  • Pair with senior colleagues and grow through structured feedback.
  • Document what you do so the team learns with you.

Required knowledge

  • Access control principles and practices
  • Systems security engineering (SSE) principles and practices
  • Adversarial machine-learning and LLM threat models including MITRE ATLAS tactics and the OWASP Top 10 for LLM Applications
  • Code tailoring tools and techniques
  • Data visualization tools and techniques
  • Supplier assessment criteria
  • Trustworthiness principles
  • Target audience requirements

Skills

  • Designing secure test plans
  • Identifying hidden patterns or relationships
  • Assessing security systems designs
  • Applying secure coding techniques
  • Performing root cause analysis

How to use this template

Copy the draft, replace every bracketed line, cut statements that do not apply to your opening, and add your compensation range and location policy. The statement lists come from the NIST NICE Framework v2.2.0, so candidates can be assessed against the same statements with Aramis:Insight.

AI Security Engineer: Mid-level

We are hiring a mid-level AI Security Engineer. You secure models, pipelines, and the applications built on them. You own this work day to day and work closely with the rest of the security team.

[Company] provides [what you do]. The security function is [size and shape: e.g., a five-person team reporting to the CTO]. Replace this paragraph with your own.

Responsibilities

  • Develop secure code and error handling
  • Document software attack surface elements
  • Identify supply chain risks for critical system elements
  • Develop software system testing and validation procedures
  • Analyze organizational cybersecurity posture trends
  • Develop system security posture trend reports
  • Develop cybersecurity supply chain risk management policy
  • Validate that procured systems and system components are genuine and unaltered
  • Detect and mitigate model-extraction attacks against a deployed prediction API, including rate, query-pattern, and output-fidelity controls
  • Construct evasion (adversarial-example) tests against a production classifier and recommend hardening measures
  • Design a prompt-injection and jailbreak test plan for an LLM application, including direct and indirect injection vectors
  • Remediate insecure output handling in an LLM feature that can generate or execute code, enforcing validation and sandboxing

Required knowledge

  • Access control principles and practices
  • Systems security engineering (SSE) principles and practices
  • Adversarial machine-learning and LLM threat models including MITRE ATLAS tactics and the OWASP Top 10 for LLM Applications
  • Code tailoring tools and techniques
  • Data visualization tools and techniques
  • Supplier assessment criteria
  • Trustworthiness principles
  • Target audience requirements
  • Cognitive biases
  • Appropriate use policies and procedures
  • Stakeholder management
  • Contracts

Skills

  • Designing secure test plans
  • Identifying hidden patterns or relationships
  • Assessing security systems designs
  • Applying secure coding techniques
  • Performing root cause analysis
  • Tailoring code analysis
  • Applying standards
  • Communicating complex concepts

How to use this template

Copy the draft, replace every bracketed line, cut statements that do not apply to your opening, and add your compensation range and location policy. The statement lists come from the NIST NICE Framework v2.2.0, so candidates can be assessed against the same statements with Aramis:Insight.

AI Security Engineer: Senior

We are hiring a senior AI Security Engineer. You secure models, pipelines, and the applications built on them. You set the direction for this work, mentor the people who do it, and answer for its results.

[Company] provides [what you do]. The security function is [size and shape: e.g., a five-person team reporting to the CTO]. Replace this paragraph with your own.

Responsibilities

  • Develop secure code and error handling
  • Document software attack surface elements
  • Identify supply chain risks for critical system elements
  • Develop software system testing and validation procedures
  • Analyze organizational cybersecurity posture trends
  • Develop system security posture trend reports
  • Develop cybersecurity supply chain risk management policy
  • Validate that procured systems and system components are genuine and unaltered
  • Detect and mitigate model-extraction attacks against a deployed prediction API, including rate, query-pattern, and output-fidelity controls
  • Construct evasion (adversarial-example) tests against a production classifier and recommend hardening measures
  • Design a prompt-injection and jailbreak test plan for an LLM application, including direct and indirect injection vectors
  • Remediate insecure output handling in an LLM feature that can generate or execute code, enforcing validation and sandboxing
  • Assess and contain retrieval-augmented-generation poisoning in a shared document or vector store
  • Scope least-privilege permissions and tool access for an agentic AI system so a compromised agent cannot exceed its intended authority
  • Adversarially test and tune a guardrail or safety-filter configuration and measure its bypass rate
  • Build AI security posture monitoring across a fleet of models and agents, inventorying models, data flows, and access
  • Mentor junior team members and review their work.
  • Represent this function to leadership and to auditors or clients.
  • Set standards, select tooling, and own the roadmap for this area.

Required knowledge

  • Access control principles and practices
  • Systems security engineering (SSE) principles and practices
  • Adversarial machine-learning and LLM threat models including MITRE ATLAS tactics and the OWASP Top 10 for LLM Applications
  • Code tailoring tools and techniques
  • Data visualization tools and techniques
  • Supplier assessment criteria
  • Trustworthiness principles
  • Target audience requirements
  • Cognitive biases
  • Appropriate use policies and procedures
  • Stakeholder management
  • Contracts
  • Contract management
  • Contractor management

Skills

  • Designing secure test plans
  • Identifying hidden patterns or relationships
  • Assessing security systems designs
  • Applying secure coding techniques
  • Performing root cause analysis
  • Tailoring code analysis
  • Applying standards
  • Communicating complex concepts
  • Communicating verbally
  • Communicating in writing

How to use this template

Copy the draft, replace every bracketed line, cut statements that do not apply to your opening, and add your compensation range and location policy. The statement lists come from the NIST NICE Framework v2.2.0, so candidates can be assessed against the same statements with Aramis:Insight.

Related roles

Roles with similar work or at a similar career stage. Each has its own job description templates and a free assessment.