Identity
WHO can access?
Solutions & introductions
Defense-in-Depth Direction — With Introductions When Helpful
This page does not assume you are buying technology. It focuses on the control patterns that reduce risk in agentic AI systems — especially in HPC research environments. When introductions to specialists genuinely help, we make them.
Why now
The Threat Model Just Stopped Being Hypothetical
In June 2026, the OWASP GenAI Security Project — an OWASP Flagship initiative with 600+ contributors across 18+ countries — released “State of Agentic AI Security and Governance v2.01,” a major update to its 2025 first edition. Its central finding: AI safety and AI security are converging at the deployment layer, and the risks once treated as speculative now have production incidents, vendor advisories, and CVEs attached to nearly every category.
This isn't abstract industry commentary. NVIDIA's Jensen Huang has been direct about the scale shift driving it: instead of roughly a billion people using computers, the next wave of computing means 100 billion autonomous agents and billions of robots using them instead — at GTC 2026 and in subsequent interviews.
The gap between those two realities is where research institutions are most exposed.
The framework
The Agentic AI Security Gap — Layer by Layer
Security and governance for agentic AI isn't one control — it's five, and most institutions have only built two or three of them:
Governance
WHAT data is appropriate to use?
Data Plane
WHERE & HOW is governed data delivered?
Runtime
HOW is the agent behaving, and what tools is it using?
Infrastructure
WHERE does the computation actually run?
The risk hiding inside this table: authorized AI can still result in inappropriate data access. An agent can have every credential it needs and still reach data it should never touch — because identity answers “who,” not “should this data be used this way.” That's the layer most environments are missing.
Authorized access does not automatically mean appropriate AI use.
The governed AI data supply chain
From Discovery to Proof — The Five Steps That Close the Gap
01
Discover
Where is the data?
02
Classify
What is it?
03
Govern
Should this AI workload be allowed to use it?
04
Deliver
Where and how should it be provided?
05
Prove
What entered the pipeline, and — just as importantly — what was excluded?
Most institutions can answer the first two. Very few can answer the last one with evidence.
Control patterns
01
One Pattern We See in Data Governance
Sensitive research data — CUI, ITAR, PHI, and “non-pattern” sensitive data that doesn't match a standard PII/PHI template — needs governance that moves with it, not governance that's only enforced where it's stored.
One architecture pattern doing this well: automated discovery and classification that goes beyond RegEx pattern-matching, paired with policy enforcement that's consistently applied across every agent-accessible pipeline — not just human-facing interfaces.
This is the model behind Secuvy's Data Security Posture Management approach, which Hammerspace selected specifically because it handles non-patterned data well. Security and governance have to be native to the data platform itself, not bolted on afterward.
02
One Pattern We See in HPC Data Architecture
Most enterprises stitch together 15 or more disconnected tools just to prepare data for AI consumption — and roughly 80% of AI project spend happens before a single dollar of ROI shows up. That's the fragmentation problem, not a compute problem.
Hammerspace's answer is a data-in-place architecture: a global namespace that unifies data across storage systems, clouds, and regions without copying it, so GPU-intensive AI applications can pull data from wherever it actually lives.
Vanderbilt University's Advanced Computing Center for Research and Education selected Hammerspace in November 2025 to modernize a 10-petabyte research environment with roughly 750 compute nodes and 80 GPU nodes. The result, per Hammerspace's announcement: a 48% reduction in average storage cost, with faster, more flexible access for the research community.
Hammerspace's composable, open architecture let Vanderbilt unify GPU server-local, tiered, and archival storage into a single environment — cutting costs while changing how compute and storage services are delivered to researchers.
03
One Pattern We See in Capacity-Tier Storage — Proven, Then Productized
Underneath ACCRE's Hammerspace deployment sits LStore — Vanderbilt's distributed, erasure-coded capacity storage layer, built to run on commodity hardware with no appliance lock-in.
Vanderbilt's cost modeling places legacy platforms at roughly $180/TB/year, while LStore runs closer to $35/TB/year. A full-capacity expansion that would otherwise force a $700K capital request can instead be handled through incremental hardware additions at a fraction of that cost.
That combination — LStore as the capacity core, Hammerspace as the orchestration layer unifying it with faster tiers above and tape archive below — is now available to other universities. Unique Checksum LLC provides Level 1 support to institutions that license LStore, making Vanderbilt's model repeatable rather than one-off.
04
One Pattern We See in Infrastructure Procurement
Storage architecture and governance don't matter if the underlying compute never gets delivered, configured, and supported correctly. Vanderbilt's GPUs, servers, and Hammerspace deployment all came through SourceCode, a technology partner founded in 1992 and headquartered in Norwood, Massachusetts.
SourceCode names Research & Education as one of its core industry verticals and lists both LStore and Hammerspace directly — the combination deployed at Vanderbilt.
For Federal research environments specifically: SourceCode is ISO 9001:2015 certified, and Comark, a SourceCode company, is an ITAR-compliant registered manufacturer. Their build process includes an in-house, U.S.-based environmental testing lab and hand-inspection at multiple stages before systems ship.
Important framing
Trusted Advisor, Not a Reseller — Here's Exactly What That Means
- We operate as a Manufacturer's Representative / Referral Partner to the vendors we introduce — a standard, well-established role in IT infrastructure sales.
- We have no authority to quote pricing or make commitments on behalf of any vendor.
- We are paid a referral fee only after a sale closes and the vendor has been paid — drawn from the vendor's own marketing budget, not added to what you pay.
- The actual seller of record for Vanderbilt's Hammerspace deployment, GPUs, and server infrastructure was SourceCode — we do not sell, ship, or invoice hardware or software ourselves.
- Our only job is to reduce confusion and get you to the right specialist conversation faster.
Start a conversation
Describe Your Scenario — We'll Point You in the Right Direction
Share the situation you're working through, and I’ll help identify the most useful next conversation.