[01]
AI is not on the asset register
Models, APIs and agentic systems get stood up outside the change process. A control framework that never enumerated them cannot tell you what they touch or who owns them.
AI Security Assessment
Know where your AI is actually exposed
Structured assessment across five pillars, scored against a versioned rubric and cited to recognized standards. Every finding gets an owner and a tracked remediation task, so the second review can be measured against the first.
- 5
- Assessment pillars
- 10
- Standards referenced
- 100%
- Findings with an owner
Every question cited to a recognized standard
- CSA AICM
- NIST AI RMF
- MITRE ATLAS
- OWASP LLM TOP 10
- ISO 27001
- NIST CSF 2.0
- EU AI ACT
The gap
Why conventional security reviews miss AI
Existing frameworks were written for systems that do not improvise. The controls still matter, but on their own they do not describe how an organization builds, buys and operates AI.
[02]
Agents hold authority nobody scoped
Delegated tool access grows quietly. Permission boundaries go untested, kill switches go unexercised, and the blast radius is discovered during the incident rather than before it.
[03]
The review is not repeatable
A point-in-time report ages out within a quarter and the next one uses a different rubric. Nothing can be compared, so nothing can be shown to have improved.
The Deepwater framework
Five pillars, one score you can compare
What we are securing
AI Landscape
Inventory of every model, API and agentic system, with sensitivity classified, provenance documented and a named business owner.
Where you stand today
Readiness
Governance and risk ownership, identity and access, vendor risk, and whether AI risk reaches the enterprise register in business terms.
Protecting the asset
Model Security
Model scanning, runtime protection, lifecycle traceability and red teaming, with secure development practices applied to AI specifically.
Security in practice
Operations
AI-aware detection and response, output monitoring for drift and manipulation, and incident playbooks distinct from conventional IR.
Building to last
Resilience
Data protection, infrastructure security and recovery planning, so the AI estate withstands disruption and continuity holds.
Each pillar scores independently and rolls into a single index. Engagements lock to a versioned question set, so a score from a year ago still means the same thing.
Read the framework →How it runs
A review you can run twice
Most AI security reviews end in a document. This one ends in a score, an owner for every finding, and a plan you can measure against next time.
How it works- [01]
Scope
A pre-flight establishes data sensitivity, system connections and agency, external providers and deployment architecture. It sets the depth of everything after it.
- [02]
Assess
The Deepwater catalog across five pillars. Every question cites a standard; every answer carries a status, an owner and evidence.
- [03]
Score
Deterministic scoring per pillar and overall, locked to a versioned question set so the number does not drift when the catalog changes.
- [04]
Remediate
Findings become owned tasks on the Deepcurrent platform, so progress is visible between assessments rather than rediscovered at the next one.
Questions
What people ask first
It depends on what the pre-flight surfaces. A focused estate with a handful of models is a very different engagement from an organization running agentic systems across several business units, and we scope it before quoting rather than after.
Find out what your AI risk actually looks like
A short call is enough to scope an assessment and tell you whether we are the right fit. If we are not, we will say so.