Strategic Data Command Inc. stratcmd.com

Security architecture and testing for systems that act on their own.

Strategic Data Command Inc. is a security consultancy. We are behind SASTER, a published 31-pattern attack taxonomy for agentic AI, and saster-harness, its open-source detection framework. We help organizations design, test, and defend the systems they run — including the AI agents they are now deploying — and we independently evaluate the fast-growing market of agent security tools sold to protect them. We also teach security teams to configure and host local open-source models for their own testing and operations, so sensitive work never has to leave their environment — drawing on twenty years across cloud, application, and infrastructure security.

2008
CISSP certified — held continuously since 2008
20+ yrs
Security architecture, pen testing & security engineering
31 patterns
SASTER taxonomy of agentic AI attacks — open source, MIT
6 venues
Research presented — DEF CON, LayerOne, GrrCON & more
01 — PRACTICES

Six practices, one standard of evidence

Every engagement is scoped in writing, tested against real systems rather than slideware, and reported in terms a leadership team and an engineering team can both act on.

PRACTICE 01

Security Architecture

Design review and architecture for the systems you are about to build — or the ones already in production that nobody has ever threat-modeled.

  • Cloud architecture review — AWS and Azure, including Azure Government
  • Threat modeling with MAESTRO and STRIDE, tied to concrete controls
  • Identity architecture — OAuth / JWT, workload identity, non-human identity governance
  • Kubernetes and service-mesh security design (RBAC, mTLS, network policy)
  • Security audit, monitoring, and logging architecture

Representative work: lead security architect for an AWS machine-learning platform at Cisco; Azure Government infrastructure security for RTX; statewide SOC and identity architecture for the California Community Colleges Technology Center.

PRACTICE 02

DevSecOps

Security that ships at the speed of your pipeline — scanning, triage, and remediation wired into the build instead of bolted on after release.

  • CI/CD pipeline security — SAST, SCA, and secrets scanning in Jenkins and modern pipelines
  • Application security posture management (ASPM) program setup and tuning
  • Shared scanning libraries and rulesets teams actually keep using
  • LLM-assisted finding triage and remediation guidance over your existing data
  • Supply-chain and dependency risk scanning for your repositories

Representative work: owned ArmorCode ASPM configuration and application onboarding at The Walt Disney Company across Checkmarx, Snyk, Edgescan, and Semgrep; built a Bedrock-based RAG pipeline that turns application findings into remediation guidance.

PRACTICE 03

Penetration Testing

Hands-on adversarial testing of networks, applications, APIs, and cloud deployments — scoped, authorized, and followed through to verified remediation.

  • Web application and API testing — REST and GraphQL
  • Cloud penetration testing — AWS and Azure environments
  • Active Directory and internal network testing
  • Mobile application testing — iOS and Android
  • Third-party test management — scoping, remediation tracking, and retest

Representative work: penetration tester for Verizon client engagements and for Cisco's internal research team; AWS and Azure testing across a 23-organization community college system; forensic follow-up on real compromises, including JBoss deserialization intrusions.

PRACTICE 04

AI Security Engineering

Security for LLM and agentic AI systems — where the failure mode is not a crashed server but an agent that did exactly what it was told, by the wrong party.

  • Agentic AI red teaming — multi-turn adversarial campaigns against your agents
  • Agent runtime detection — behavioral baselining and drift detection with saster-harness
  • Guard-model and filter evaluation, including side-channel extraction testing
  • MCP server and agent-to-tool trust-boundary assessment
  • Agentic identity architecture — workload identity and human-in-the-loop controls for high-risk actions
  • AI threat modeling for teams building with LLMs

Representative work: author of the SASTER taxonomy and saster-harness; SPIFFE/SPIRE workload-identity design for an agentic AWS deployment; Bedrock/Claude alert triage inside an OpenSearch-based SOC.

PRACTICE 05 — NEW

AI Agent Security Tool Evaluations

Right now it’s the wild west. Dozens of new agent security products — guardrails, prompt-injection detectors, agent firewalls, MCP scanners, runtime monitors — are shipping with bold detection claims, and there is no reliable, independent evaluation to check them against. Vendor demos test the attacks the vendor chose. We test the ones they didn’t. For buyers, that means a shortlist grounded in evidence before you sign; for vendors, independent validation you can put in front of a skeptical customer.

  • Head-to-head evaluation of agent security tools — guard models, filters, firewalls, scanners, and runtime detection — in your environment, on your agents and models
  • Coverage mapped to the SASTER 31 patterns: what each tool actually catches, what it misses, and where its claims are specification-only
  • Bypass & evasion testing — multi-turn campaigns, indirect injection, and sequences of individually permitted tool calls, not just single-prompt test sets
  • False-positive and false-negative measurement against realistic benign workloads, so a tool that blocks everything doesn’t score as “secure”
  • Blind-spot analysis across the tool, model, and process planes — including state-based outcomes that action-level allow-lists and dashboards stay green through
  • Operational fit: latency, cost per call, deployment model, telemetry quality, and how alerts land in your existing SOC and SIEM

What you get: a written scorecard per tool — SASTER coverage, bypass results with reproduction detail, false-positive rates, and a buy / don’t-buy / buy-with-conditions recommendation you can take into procurement. No vendor sponsorship, no paid placement: we work for the buyer, or for a vendor willing to be tested on the same terms as everyone else.

Why us: this is the work behind our published research — guard-model evaluation including side-channel extraction through a guard LLM’s own refusal signal (DEF CON 34 AI Village), live demonstrations at GrrCON and LayerOne where five layered defenses stayed green through a boundary crossing, and the open saster-harness detector baseline we measure other tools against. It also isn’t new territory: Larry’s comparative evaluations of web application vulnerability scanners are widely cited and predate the AI tool wave by over a decade.

PRACTICE 06 — NEW

Local Open-Source Models for Security Testing & Operations

Some security work should never go to a third-party API — exploit code you’re analyzing, incident data, internal logs, proprietary source code, client findings. We are experts at teaching companies to run the alternative themselves: open-source, open-weight models hosted locally, on hardware and infrastructure you control. This is hands-on enablement, not a slide deck — we configure the stack with your team, then teach them to run, secure, and improve it after we leave.

  • Model selection for security work — matching open-source models to the task (code review, log and alert triage, threat-intel summarization, detection engineering, pen-test assistance) and testing them on your data before you commit
  • Local hosting & configuration — GPU sizing, inference serving (for example vLLM, Ollama, and llama.cpp-class runtimes), quantization trade-offs, context and performance tuning, on-prem, private-cloud, and air-gapped deployments
  • Secure-by-design setup — network isolation, access control and authentication in front of the model, prompt and output logging that doesn’t itself leak sensitive data, and secrets handling
  • Workflows for security testing — local models wired into pen-testing, code and configuration review, and red-team preparation, with guardrails and human review where the output drives an action
  • Workflows for security operations — SOC triage, incident enrichment, and reporting assistance running entirely inside your environment, integrated with your existing tools and SIEM
  • Hands-on training for your team — live workshops and paired build sessions, plus runbooks covering model provenance and licensing checks, updates, evaluation, and day-to-day operations

What you get: a working local model deployment configured for your security testing and operations use cases, a hardened reference configuration your team can reproduce, and trained staff who can operate, evaluate, and extend it — no dependency on us, and no sensitive prompts, code, or incident data sent to an outside model provider to get there.

Why us: the same team behind SASTER and saster-harness — we run local models in our own research and detection work, evaluate guard models and agent tooling for a living, and bring twenty years of cloud, infrastructure, and SOC architecture to the hosting side: sizing it, isolating it, and logging it like any other security-critical system.

02 — OPEN RESEARCH

The AI practice runs on published research

The taxonomy and detection tooling behind our AI security work are open source. You can read the code, check the claims, and see exactly where coverage ends — partial coverage is stated plainly rather than left to be inferred.

SASTER — attack taxonomy

A 31-pattern adversarial taxonomy for agentic AI systems in four tiers, from direct prompt injection through epistemic exploitation and compositional capability emergence. It turns abstract AI risk categories into specific, testable threats — and ships with an OWASP GenAI mapping matrix and detector determinism documentation.

saster-harness — detection framework

An open-source agentic detection and response framework in Python: an instrumentation layer between an agent and its tools, with per-session behavioral baselining and 13 detectors across 9 attack patterns. It instruments the tool, model, and process planes — surfacing agent actions that never appear as a tool call and that conventional telemetry and DLP are structurally blind to.

The finding underneath the tooling

In live demonstrations presented at GrrCON and LayerOne, an agent with a legitimate job, legitimate tools, and a written boundary crossed that boundary through a sequence of individually permitted actions. Action-level allow-lists approved every step; five defenses stayed green. Only an evaluator reading world state afterwards caught it. Our conclusion: containment for agents has to be measured on state, by an evaluator that shares no input with the agent, and enforced in infrastructure the agent cannot rewrite. That principle drives every AI engagement we run.

Explore the taxonomy

SASTER organizes 31 patterns into four tiers of increasing subtlety. Filter by tier to see what each level of attack looks like.

Representative patterns shown. The full 31-pattern specification — including which patterns have a reference detector in saster-harness and which are specification-only — is published in the SASTER repository.

03 — SPEAKING

Research presentations

Slides, recordings, the taxonomy document, or an engagement reference are available on request.

GrrCON 2026

Your AI Agent Could Be a Threat Actor

Live bypass of a multi-layer agent defense stack — and continuous measurement of where those defenses hold.

DEF CON 34

AI Village — research poster

Block Means Yes: extracting protected agent data through a guard LLM's own refusal signal.

LayerOne 2026

Hunting Rogue Autonomous Agents

A field guide to AI threats the SOC can't see.

HushCon 2025

When Agents Attack

A2A adversarial collaboration in practice.

CactusCon 2024

Building a Backyard Threat Intelligence Platform

That works for you.

DerbyCon 2019

kubered

Recipes for C2 operations on Kubernetes.

04 — ENGAGEMENTS

How an engagement runs

STEP 1

Scope & authorization

Targets, rules of engagement, and exclusions are agreed in writing before any testing begins. No written authorization, no testing.

STEP 2

Assess

Architecture review, threat modeling, or hands-on testing against the agreed scope — with findings captured as evidence, not impressions.

STEP 3

Report

Findings ranked by real-world impact, each with reproduction detail and a remediation path your engineers can execute.

STEP 4

Retest & verify

Fixes are retested to confirm they hold. For AI systems, that means verifying detection actually fires — not assuming it does.

AUTHORIZED WORK ONLY

Adversarial testing has a bright line, and we stay on the right side of it

All offensive and adversarial work — penetration testing, agent red teaming, detection evaluation — is performed only against systems we own, systems covered by a written engagement with their operator, or public testing platforms whose published terms authorize the testing. Scope is enforced in our tooling, not just by convention: saster-harness, for example, deploys only against an explicit allowlist of hosts.

05 — ABOUT

Larry Suto — Co-Founder, CTO & Principal Consultant

Larry Suto is Co-Founder, CTO, and Principal Consultant at Strategic Data Command Inc. — a security architect and independent AI security researcher whose current work focuses on the runtime behavior of autonomous agents — what they actually do once deployed, as opposed to what their designers intended. He is the author of SASTER, a published 31-pattern attack taxonomy for agentic AI systems, and of saster-harness, its open-source detection counterpart.

Underneath the AI research sit twenty years of application security, penetration testing, and cloud security architecture: security architecture and pen testing at Cisco, application security at The Walt Disney Company, Azure Government engineering support for RTX, and security operations for a 23-organization community college system in California. His earlier published research — comparative evaluations of web application vulnerability scanners — is widely cited and predates the AI work by over a decade.

Strategic Data Command Inc. is based in Pine, Arizona, and works with clients remotely across the United States.

ROLES & ENGAGEMENTS HAVE INCLUDED

The Walt Disney CompanyCisco SystemsRTX California Community Colleges Tech CenterG-PVerizon UC Irvine HealthWells FargoKaiser PermanentePepsiCo

Tell us what you're building — or what you're worried about.

Your point of contact is Larry Suto, Co-Founder, CTO & Principal Consultant. Engagement inquiries, tool evaluation requests, local open-source model training and hosting inquiries, speaking invitations, and questions about the open-source research are all welcome — reach him directly at the email and phone below. If your request involves testing, include who owns the target system — it speeds up scoping.

Larry Suto — Co-Founder, CTO & Principal Consultant Strategic Data Command Inc. Pine, Arizona 415-999-0925 x.com/tlbs1970 stratcmd.com