TECHNOLOGY · SECURITY · ADVISORY
We help organizations design, secure, engineer and transform technology across applications, AI, infrastructure, compliance and digital platforms.
01WHAT WE ARE
One practice across the full life of a system. We design and build technology, we assess and secure it, and we help you govern it well enough to show a regulator, an auditor or your largest customer.
We design and develop software, applications, digital platforms, SaaS products, AI solutions and technology-enabled services.
We assess, engineer and secure applications, AI systems, cloud environments, infrastructure and digital technologies.
We help organizations navigate security architecture, privacy, GRC, ISO, compliance, risk and technology transformation.
85 services across assessment, engineering, build, governance and advisory. Most engagements draw on two or three.
Every component moves trust somewhere new, and every label under it is a class of finding we test for. This is the scope of an AI security assessment.
USER
Untrusted input, always
AI APPLICATION
Where policy is enforced: or is not
LLM
Instruction-following, not judgement
RAG
Retrieved content is executable text
AGENT
Plans, loops and acts on your behalf
TOOLS
Every tool is a new privilege
MCP
Transitive trust across servers
ENTERPRISE DATA
The boundary that must hold
We do not hand over a list of issues. We show the chain, how one weak assumption becomes access to the thing you actually care about, and what breaking the chain costs.
Entry
Attacker
Unauthenticated, internet-based
No credentials, no prior access, no insider. Everything that follows is earned from a standing start, which is what makes the chain worth taking seriously.
Exposure
Public API
Undocumented v1 route still reachable
A deprecated endpoint nobody owns is still routed, still authenticating, and absent from the asset inventory the security team works from.
Vulnerability
Broken authorization
Object ID accepted without ownership check
Any authenticated user can read any object by changing an identifier. Individually this scores medium. In this chain it is the hinge everything turns on.
Pivot
Internal service
Trusted network call, no second check
The internal service trusts the caller because the call came from inside the perimeter. The perimeter stopped being a boundary two steps ago.
Privilege
Cloud role
Service identity scoped far too widely
The workload's role can read every bucket in the account rather than the one it needs. Convenience at provisioning time becomes blast radius at breach time.
Objective
Crown jewel
Customer data store, fully readable
The full customer dataset, readable and exfiltratable without a single alert firing. This is the number that appears in the breach notification.
One continuous engagement model. Each stage hands the next something real, and the last one hands you evidence.
Establish the real picture
Systems, exposure, ownership and business impact, mapped as they actually are rather than as the architecture diagram claims. Nothing useful follows from a scope built on assumptions.
Turn findings into decisions
Findings ranked by consequence rather than severity label, with trade-offs made explicit. Leadership gets a small number of decisions to make, each with a cost and a reason.
Design what closes the gap
Reference architectures, control design and the operating model that keeps them working after we leave. Design-level fixes close classes of issues rather than individual tickets.
Prove it holds
Validation that the change worked, documented to a standard that survives a customer security review, a regulator or an audit committee.
Security problems rhyme across sectors, but the consequences do not. These are the environments we know well enough to be useful quickly.
Anonymised, but real. Each ran the same way: assess the actual exposure, advise on the decisions that matter, design the fix, prove it holds.
89 findings
misconfigurations closed across AWS and Azure
Adversarial
vulnerabilities found in a clinical ML model
2 hours
to containment, 100% of data recovered
Four things the engagements keep proving. Recognising them on day one is most of what makes an assessment fast.
Agent & MCP trust
An agent's real authority is always wider than its operator believes.
Permission scope compounds across tool servers. We map effective authority rather than declared authority, because the gap between the two is where agent compromise actually happens.
ASSESSED ON EVERY AI ENGAGEMENT
Retrieval integrity
Retrieved content is executable text, and most pipelines treat it as data.
We test RAG against an attacker who can write to the corpus, because that is the only threat model that matters once retrieval sits in the loop. Provenance and isolation hold. Output filtering does not.
STANDARD SCOPE IN LLM TESTING
Cloud privilege paths
The route from build pipeline to production data is usually three steps.
The same role-assumption chains recur across estates because the same provisioning habits do. We look for them first, which is why cloud assessments close faster than teams expect.
FOUND IN MOST MULTI-ACCOUNT ESTATES
Assurance & evidence
Most penetration-test reports fail a customer security review.
Auditors, enterprise buyers and regulators each accept different evidence. We write to all three from the first day of an engagement, which is why our reports survive procurement instead of triggering another round.
BUILT INTO EVERY DELIVERABLE
Published analysis
Findings are mapped to the frameworks your regulators, customers and certification bodies already use, so the work lands as evidence rather than opinion.
EU AI Act
AI regulatory readiness
NIST AI RMF
AI risk management
ISO/IEC 42001
AI management systems
ISO 27001
Information security management
NIST CSF
Program benchmarking
NIST 800-53
Control baselines
GDPR
Privacy and data protection
OWASP LLM Top 10
AI failure classes
If yours is not here, ask it directly: the first reply comes from a practitioner, not a sales desk.
Cybaethrex Training
Cybaethrex Training delivers professional, corporate and university programs in AI security, cloud security and offensive security, taught by the same practitioners who run the engagements.
Next step
Tell us what you are building, adopting or worried about. You will get a practitioner’s view of where the risk actually sits, before any engagement is scoped.