Governance · AI Risk

Assumed Risk:
The Vulnerability
Organisations Normalise

In most organisations, the biggest security weakness isn't a misconfigured firewall or an exposed system. It's something far less visible — assumed risk. And in GenAI environments, it becomes a force multiplier.

KS
Kevin Stout
Founder, OneCompliant s.r.o. · AI Security Executive
6 min read
GovernanceAI Risk
Assumed risk in AI governance

In most organisations, the biggest security weakness isn't a misconfigured firewall or an exposed system. It's something far less visible: assumed risk.

This is the risk no one formally accepts. The risk that lives in statements like:

"We'll fix it later"
"It's internal only"
"The model endpoint isn't discoverable"
"That service account is just for automation"
"The vendor says it's secure"
"Governance will come after the AI rollout"

These aren't controls. They're assumptions.

What is assumed risk?

It's the gap between what the organisation believes is controlled versus what is actually secured, monitored, and enforceable.

Unlike formal risk, assumed risk is undocumented, unowned, and unvalidated. It is often invisible across teams. And it spreads quietly.

Why GenAI changes everything

Generative AI doesn't operate in a contained system. It depends on data pipelines, APIs and orchestration layers, external prompts and inputs, vector stores and embeddings, third-party models and plugins, and continuous retraining and feedback loops.

This creates a massive, interconnected attack surface. Now combine that with assumed risk.

Assumptions become attack paths.

What this looks like in reality

Overprivileged service accounts controlling entire pipelinesAccounts created for "just automation" accumulate permissions over time and become high-value targets with broad blast radius.
Model APIs exposed without abuse controlsEndpoints deployed for internal use that gradually become accessible — without rate limiting, authentication hardening, or monitoring.
Prompt logs retained indefinitely without governanceInteraction logs that contain sensitive user data, internal context, and business logic — retained without classification, access controls, or retention policies.
External data trusted without validationRAG systems and AI pipelines that ingest external content without source trust controls, content classification, or prompt injection detection.
"Temporary" exceptions that become permanent architectureControls bypassed for speed during pilots that are never reinstated — because the system is now in production and nobody wants to disrupt it.
Policies that exist but are never enforcedAI governance documents that describe expected behaviour but have no technical implementation, no monitoring, and no accountability.

In a GenAI environment, small gaps don't stay small. They scale.

The real risk

This isn't just technical exposure. It leads to data leakage across systems, prompt injection and model manipulation, loss of auditability, regulatory exposure, and brand and trust damage.

And most importantly: loss of control over how your AI behaves.

Why it keeps happening

Because it's easy to justify. Speed over discipline. Innovation over governance. Shared ownership — which becomes no ownership. Pilot-mode thinking carried into production.

None of these are deliberate failures. They are the predictable result of deploying AI faster than governance can follow.

What CISOs and leaders must do

Stop assuming risk. Start governing it.

Turn assumptions into explicit, owned decisions — every "we'll fix it later" needs an owner, a timeline, and a formal risk acceptance if it stays.
Enforce controls through architecture, not policy alone — if a control exists only in a document, it doesn't exist in production.
Apply least privilege across identities, pipelines, and models — not just user accounts. Service identities, model credentials, and pipeline permissions all apply.
Treat data lineage as a security control — know where data enters your AI systems, how it flows, and where it can exit.
Build AI-specific detection and response — your existing SIEM rules were not written for prompt injection, context poisoning, or semantic data extraction.
Create cross-functional accountability — AI security is not a CISO problem alone. Security, AI, data, legal, and product teams all own a part of it.

A better question to ask

Not: "Is this secure?"

But: "What are we assuming — and what happens if we're wrong?"

Final thought

In cloud environments, assumed risk is dangerous.

In GenAI environments, it becomes a force multiplier. Because assumptions don't stay isolated. They compound.

Assumed risk isn't just a security gap. In GenAI, it's a strategic vulnerability.

#CISO#AI#GenAI#AISecurity#AIGovernance#RiskManagement#OneCompliant

Turn your assumed risk into governed risk

OneCompliant's OASAT assessment surfaces the gaps between what your organisation believes is controlled and what is actually secured — across your AI systems, pipelines, and controls.

Book OASAT AssessmentExplore the Platform