AI security and governance · built by Data Science and Engineering Experts NAICS 541330 / 541511 / 541512 / 541519  ·  CMMC-aware
§00·AI Security and Governance·built by Data Science and Engineering Experts
senior-only bench · operator-led · OWASP LLM Top 10 · MITRE ATLAS · NIST AI RMF

Your AI has attack surfaces
your security team
hasn't mapped yet. We find them, document them, and build the governance controls to close them. Senior-led AI security testing and governance readiness, mapped to the OWASP LLM Top 10, MITRE ATLAS, and the NIST AI RMF. The person who scopes the work is the person on the keyboard.

methodology
OWASP
LLM Top 10 (2025) coverage
adversarial
ATLAS
MITRE technique mapping
governance
NIST
AI RMF readiness mapping
authored IP
mcp-warden
open-source, MIT, signed releases
delivery
senior-only
principal on the keyboard
entry
2wk
fixed-fee X-Ray, first findings in 48h
§ Start here·which problem is louder?

Start here. Which problem is louder?

Most teams come to us mid-problem, not shopping for a service. Pick the line that sounds like your quarter, and we will point you at the fix, the proof it works, and something worth reading on the way.

fig 01·what we run, end-to-end·representative of how we ship
Open source ↗
~/dsee   cat /var/log/shipping.sample illustrative
[gate] check ok mcp-warden warden.lock · drift gate · CI
[eval] eval pass rag/v3 golden suite green · drift in band
[deploy] deploy ok fedgov/ingest sam.gov sync · batch loaded
[scan] scan bedrock-iam clean ✓
[release]release mcp/gemini-bridge tagged · github
[deploy] deploy ok dsealgo risk-circuit · zero downtime
[review] red-team client-X findings triaged · prioritized
[release]release mcp/perplexity-async tagged · github
[deploy] deploy ok foodee/payments stripe · PII-clean
_
an illustrative slice of the kinds of work we ship Open source ↗
What we run, end-to-end.
fig 01 · stack
Auth / Identity
Clerk · JWT · OIDC
least-priv IAM
API Gateway
66 routes
p95 < 200ms
LLM Routing
LiteLLM · 5 providers
cost-routed
Retrieval
pgvector · BM25+dense
hybrid search
Evals
CI gates · golden suite
drift-tracked
Inference
Bedrock · vLLM
GPU-shared
Observability
Traces · logs · cost
per-tenant
AI Security
Red-team · STRIDE/AI
NIST AI RMF
Governance
EU AI Act · ISO 42001
CMMC-ready
every cell = a thing we'll write, run, and document for you. AWS-native
§ Why us.
Real IP, not slideware.

We don't just test AI security. We build it.

The tooling we deploy in an assessment is tooling we wrote and open-sourced. The posture we bring to your system is the posture we use on our own.

authored IP

mcp-warden

A principal authored it. An open-source MCP supply-chain lockfile and CI gate that pins an MCP server's declared tool surface into a signed lock and fails CI when that surface drifts. The kind of tooling we deploy in assessments. MIT licensed, on PyPI as mcp-warden-cli, signed releases, 400+ tests.

Inspect mcp-warden on GitHub ↗
also ours

conclave

A multi-model adversarial council we built to pressure-test our own designs. It is the same posture we bring to your AI system: assume the model is hostile and prove otherwise.

how we staff

Senior-only bench

Engagements run on a published method, the OWASP LLM Top 10 and MITRE ATLAS, by senior practitioners. No junior hand-off, no rented dashboard. The person who scopes is the person on the keyboard.

§01·Secure your AI·find what is exploitable

Find prompt injection, data leakage, and agent abuse before attackers do.

We red-team the whole stack an attacker sees, not just the prompt box. Then we hand you evidence-backed findings and a remediation roadmap, mapped to the OWASP LLM Top 10 and MITRE ATLAS. Fixed-fee, fixed-scope, principal-run.

primary

AI Security X-Ray

2 weeks · $12k to $18k fixed

A point-in-time threat model and adversarial test of one AI system, with severity-ranked findings, remediation, and a runbook. First findings inside 48 hours.

Scope the X-Ray
early access · limited availability · scoping call required
deeper engagement

AI Red Team Sprint

4 weeks · indicative $35k to $55k

A full adversarial campaign across the system, multi-turn attacks, chained exploits, and agent steering, when a two-week X-Ray is not enough.

Request a scoping call
what comes after

AI Security Co-Pilot

retainer · from $8.5k/mo

Ongoing security oversight once the system is live, keeping the red-team harness, the findings backlog, and the AI inventory current as the system changes.

The five attack surfaces we test
  1. Input and output. Direct and indirect prompt injection, jailbreaks, multi-turn Crescendo, system-prompt leakage, improper output handling. LLM01 · LLM02 · LLM05 · LLM07
  2. Retrieval (RAG / vector DB). RAG poisoning, embedding-space attacks, unauthorized document retrieval, context-window disclosure. LLM02 · LLM08
  3. Tool and agentic layer. Excessive agency, tool abuse, confused-deputy chains. Can the agent be steered out of its purpose, and can you stop it. LLM06 · Agentic Top 10
  4. Model and supply chain. Provenance, poisoning exposure, MCP supply-chain review using the integrity checks shipped in mcp-warden. LLM03 · LLM04
  5. Runtime and ops. Unbounded consumption and cost-amplification, guardrail bypass, logging gaps, missing rate, spend, and abuse controls. LLM10
The method, five phases: Scope and threat-model → recon and architecture review → exploitation → verification and risk scoring → report and fix plan. Anchored to the OWASP LLM Top 10 and MITRE ATLAS.
See the full AI Security Assessment
§02·Govern your AI·prove it is controlled

Governance your auditors, regulators, and buyers will accept.

A fixed-fee readiness engagement for teams putting AI into production. We inventory your AI, classify risk, find the gaps against the framework you care about, the NIST AI RMF, the EU AI Act, or ISO/IEC 42001, and layer it onto the SOC 2 or ISO 27001 program you already run. Readiness and alignment, not certification.

primary

AI Governance Readiness Snapshot

4 to 6 weeks · $25k to $45k fixed

AI inventory, risk classification, and a gap assessment against one framework, crosswalked onto your existing SOC 2 or ISO 27001 program, with a prioritized remediation roadmap.

Scope the Snapshot
what comes after

Fractional AI Compliance Officer

retainer · from $7.5k/mo

Recurring ownership of the risk register, framework interpretation, and audit-ready evidence once the readiness work is done, reporting to your board and insurer.

govcon add-on

AI Risk and Compliance Plan for RFP/TO

2 to 4 weeks · $15k to $35k · often B&P

An AI risk and compliance narrative for a specific proposal or task order, mapped to NIST AI RMF, frequently funded as bid and proposal.

What readiness means here: we map your AI to the framework you care about and align it onto the controls you already run. We produce readiness evidence and a remediation roadmap. We do not certify, guarantee, or attest compliance, that is your auditor's role, and we set you up to pass it.
See AI Governance Readiness
§03 Why DSE.
Boutique posture.

A small firm of senior practitioners, specialized in the AI attack surface.

We are not a pyramid. There is no junior hand-off, no rented dashboard, and no thesis to push. We pick a narrow problem, the security and governance of the AI you are actually shipping, and we go deep.

published method

OWASP, ATLAS, NIST AI RMF.

Coverage is organized against the OWASP LLM Top 10 and MITRE ATLAS, and governance is mapped to the NIST AI RMF. You can audit the method, not just trust it.

authored IP

mcp-warden and conclave.

mcp-warden is open-source, MIT, with signed releases, on PyPI as mcp-warden-cli. conclave is the adversarial council we built to pressure-test our own designs.

specific surfaces

We tell you what we test.

The five attack surfaces are enumerated above, input and output, retrieval, the tool and agentic layer, the model and supply chain, and runtime and ops. You know exactly what we test before you pay.

senior-led scoping

Scoper equals deliverer.

The person who runs the senior-led scoping call is the person who delivers the work. No hand-off to a team you never met.

deep technical writing

We work in the open.

We publish how we test, in detail, including a walkthrough of how we run an OWASP LLM Top 10 assessment.

ReadHow we test the OWASP LLM Top 10
Data Science & Engineering Experts came up building production AI systems, RAG pipelines, agents, multi-tenant platforms, so when we secure and govern AI, we read the architecture the way the engineers who built it do. The data science and data engineering practice still runs; it is no longer the headline. We are the same data experts, now pointed at the AI attack surface. See our data engineering services and data science work
§05  What we won't take on

Listing what we decline
is the strongest signal
we have standards.

Nobody publishes this. We do. If you are looking for one of the engagements below, we will happily refer you elsewhere, and the rest of this site means more because of it.

If your problem isn't here, it's likely one of ours.
§04 Who we serve.
Two buyers, one bench.

Two buyers, one senior bench.

The same senior practitioners secure and govern AI for commercial teams shipping to enterprise buyers and for federal programs putting AI into production.

B2B SaaS / mid-market

Shipping AI to enterprise buyers.

Teams shipping RAG apps, copilots, and agent platforms who need to know what is exploitable and prove to enterprise buyers that the AI is governed. SOC 2 / ISO 27001 crosswalk, EU AI Act and ISO 42001 readiness, fixed-fee.

Federal

Unclassified AI governance and review.

Unclassified AI governance readiness, pre-deployment AI architecture review and threat modeling, AI/ML supply-chain security review, an AI risk register and policy, NIST AI RMF mapping, and an AI governance narrative for proposals (B&P).

Federal capability
Atlanta / regional

Local senior help, in the open.

Atlanta-based, working with regional SaaS, fintech, and the partner network (MSPs, fractional CFOs, post-funding and M&A readiness). Local senior help on the AI attack surface.

§06 Insights.
Receipts you can read.

We work in the open, and we publish how.

Most firms hide their method and their code. Ours is on GitHub and in the Refinery Report. Read why AI projects stall, what the real ROI looks like, and how we run an OWASP LLM Top 10 assessment.

github.com/ernestprovo23 · OSS Open source ↗
mcp-wardenMCP CI gateMIT ↗ conclaveadversarial councilopen ↗
mcp-warden ships signed releases on PyPI as mcp-warden-cli.
The tooling we deploy in assessments is tooling we open-source.