HTB Certified Offensive AI Expert (COAE)
Hack The Box
Credential ID: HTBCERT-1287C8C6C3
Verify credentialPortfolioAI Security
AdrianInfantes
AI Security Architect
- Hack The Box Global Top 10
- HTB COAE
- AWS Security – Specialty
- AWS Generative AI Developer – Professional
- 5 AWS architectures
About
I'm an AI Security Architect with nine years in machine learning and AI, from data science and production ML to securing agentic systems. I design guardrail platforms, protected agent runtimes, and evidence-led release paths for LLM agents and RAG copilots on AWS, and I test them the way an attacker would.
Architecture folio / Plates 01–05
AWS Architect
Five AWS architectures for controlled, observable AI systems. Explore each problem, decision, and limitation in an interactive map.
Guardrails as code
- Problem
- Teams need consistent guardrail policies and an explicit version for each application to use.
- Decision
- Define reusable profiles in Terraform and publish a versioned reference through SSM.
- Limitation
- Provisioning a policy does not prove its effectiveness or that an application applies it. Both need separate checks.
- Technical scope
- Amazon Bedrock Guardrails, Terraform, versioned SSM contract
Protected agent runtime
- Problem
- Model inputs and outputs need protection while tool permissions remain under application control.
- Decision
- Apply input and output checks around Bedrock, with deterministic authorization before tool execution.
- Limitation
- Guardrails assess content; they do not grant permission to use tools or access data.
- Technical scope
- Amazon Bedrock, Bedrock Guardrails, IAM, deterministic tool boundary
Evidence-led operations
- Problem
- AI releases need evidence for promotion and clear signals for rollback without logging sensitive payloads.
- Decision
- Combine evaluation gates and telemetry without payloads with progressive release and explicit rollback.
- Limitation
- Release quality depends on evaluation coverage and chosen thresholds. Telemetry alone cannot demonstrate safety.
- Technical scope
- Evaluation gates, CloudWatch, progressive rollout, rollback
Shared guardrail platform
- Problem
- Agent teams in separate AWS accounts need the same content controls without rebuilding them in every project.
- Decision
- One Terraform pipeline deploys a versioned guardrail to each consumer account and publishes its contract in SSM. Agents resolve it at deploy time while enforcement moves from observe to pilot to enforce.
- Limitation
- A shared baseline cannot fit every use case. Project overlays need their own evaluation, and enabling enforcement does not prove that every consumer applies it.
- Technical scope
- Amazon Bedrock Guardrails, Terraform, multi-account SSM contract, staged enforcement
Guarded RAG copilot
- Problem
- A contact-center copilot must ground its suggestions in internal knowledge and case context while the operator stays in control.
- Decision
- An intake agent prepares the case in memory. A copilot on AgentCore retrieves from a knowledge base, checks its output with Bedrock Guardrails, and sends suggestion cards to the operator through an event bus.
- Limitation
- Suggestions are advice, not actions: the operator decides. Retrieval quality and guardrail thresholds need continuous evaluation, and redacted traces trade debugging detail for privacy.
- Technical scope
- Amazon Bedrock AgentCore Runtime and Memory, knowledge-base retrieval, Bedrock Guardrails, Kafka, Langfuse
Certifications
AWS Certified Generative AI Developer – Professional
Amazon Web Services
Completed
AWS Certified Security – Specialty
Amazon Web Services
Completed
AI Red Teaming
Hack The Box Global Top 10
L4tentNoise
View Hack The Box profile
AI Security Write-ups
Four analyses of Hack The Box challenges, focused on defensive lessons and the limits of the evidence.
HTB challenge analysis and reproduction
Doctrine Studio: A Prompt Is Not a Permission Boundary (opens in a new tab)
Permission checks belong in the application. A model's instructions cannot replace them.
HTB challenge analysis and reproduction
Power Supply: Sensitive Data Stays Sensitive Across Turns (opens in a new tab)
Data keeps its sensitivity throughout a conversation. Access checks must hold across turns.
HTB challenge analysis and reproduction
Neural Detonator: Models Can Carry Code (opens in a new tab)
Treat ML artifacts as software: verify their provenance, inspect them, and isolate their execution.
Partial HTB challenge reproduction
Sigma Technology: A Partial Reproduction and Missing Evidence (opens in a new tab)
Separate validated observations from missing evidence. The original input is unavailable, so the reproduction remains partial.
Experience
BBVA
AI Security Architect
Designing AI security architectures and adversarial evaluations across models, data, tools, and human decisions.
Ecoembes
Machine Learning Engineer
Built machine-learning systems across computer vision, document NLP, route optimisation, and time-series forecasting, from data preparation to integration and monitoring.
Capgemini
Data Scientist
Developed pricing and predictive-analytics models, operational Power BI reporting, and inventory decision support.
Education
UNED
Bachelor’s Degree in Artificial Intelligence Engineering
Current undergraduate studies in artificial intelligence engineering.
MIOTI
Big Data & Data Science Master
240-hour programme in big data and data science.
MIOTI
Master in Machine Learning & Deep Learning
240-hour programme focused on machine learning and deep learning.
Contact
Available for AI security architecture and AI red-team engagements.
Message me on LinkedIn
The AI assistant on this page is a small security exercise of its own: it answers only from public facts, treats visitor messages as untrusted data, is rate-limited, and replies in plain text.