Know every AI system touching your data – before an attacker or auditor does.
Armorstack’s AI Security Assessment is a structured evaluation of models, prompts, RAG layers, APIs, plugins, non-human identities, and the infrastructure around them. We inventory sanctioned and shadow AI, pressure-test the system boundary, and score findings to NIST AI RMF and ISO/IEC 42001 – with a remediation roadmap your board can act on.
What an AI security assessment actually covers
Unlike a model-only test, we examine the full system boundary:
Models, prompts, and agents
Direct and indirect prompt injection, jailbreaks, cross-prompt injection, output hijacking, and unsafe tool use by autonomous agents (OWASP Agentic AI / LLM Top 10).
Data pipelines and lineage
Training and fine-tuning data exposure, runtime prompt leakage, PII in context windows, retrieval corpus poisoning risk, and unauthorized memorization paths.
Infrastructure and integrations
Cloud AI services (Azure OpenAI, Amazon Bedrock, OpenAI, and peers), vector databases, overprivileged APIs, MCP and plugin surfaces, GPU and workload posture where in scope.
Governance and non-human identities
Shadow AI adoption, role and service-account privilege, API key hygiene and rotation, secret leakage into prompts or logs (OWASP NHI), and compliance readiness for your regulatory stack.
How Armorstack runs it
Analyst-led assessment backed by Armorstack discovery and observability where deployed. Not a self-serve AI-SPM product pitch.
Map sanctioned tools, embedded SaaS AI, browser/extension shadow AI, and agentic workflows. CORE Fabric and Miralto feed asset, app, and egress context where deployed; analysts close the human and process gaps tooling alone misses.
Rate each use case against NIST AI RMF (Govern, Map, Measure, Manage) and ISO/IEC 42001, with explicit callouts to Agentic and Non-Human Identity failure modes.
Prioritize what is exposed, what an attacker can reach, and what an auditor will ask next – not a generic maturity slideshow.
Executive PDF, domain scores, and a sequenced remediation plan. Optional path into ongoing SENTRY observability (CORE Enclave + Miralto + Sentry Pulse / Pulse AI + Huntress) and VERITY governance.
Domains we score
- AI governance and strategy – Ownership, policy, board reporting, Acceptable Use.
- Data security and privacy – What AI can see, store, and exfiltrate (PHI, CUI, NPI, customer data).
- Threat landscape and adversarial risk – Prompt injection, jailbreak, agent tool abuse, model supply chain.
- Security observability – Whether you can see AI use the way you see endpoints and network.
- Compliance and frameworks – NIST AI RMF, ISO/IEC 42001, and sector overlays (HIPAA, CMMC, SOC 2) where they apply.
- Workforce and culture – Literacy, shadow usage, training that changes behavior.
- Infrastructure, access, and non-human identity – Keys, service principals, integrations, overprivilege.
Choose your assessment tier
Tiers are set by number of users. Your tier is confirmed on the scoping call before kickoff.
One engagement, delivered in five business days, with the same deliverables at every tier. The five-day clock starts once access and discovery inputs are ready.
Essential
- 1 primary identity tenant
- 2 admin/telemetry sources
- 3 AI apps reviewed in depth
- 2 agents/integrations sampled
- 1 stakeholder interview
- 45-minute executive review
Business
- 1 primary identity tenant
- 3 admin/telemetry sources
- 5 AI apps reviewed in depth
- 4 agents/integrations sampled
- 2 stakeholder interviews
- 45-minute executive review
Professional
- 2 primary identity tenants
- 5 admin/telemetry sources
- 8 AI apps reviewed in depth
- 8 agents/integrations sampled
- 3 stakeholder interviews
- 60-minute executive review
Enterprise
- 3 primary identity tenants
- 6 admin/telemetry sources
- 10 AI apps reviewed in depth
- 12 agents/integrations sampled
- 4 stakeholder interviews
- 60-minute executive review
Quarterly reassessments and an AI red-team deep dive are scoped separately.
Register to begin your assessment
Complete the form. We confirm scope and send access within one business day.
Prefer to talk to a person? Call 877-890-5508 or email [email protected].
FAQ
What is in scope?
Models, prompts, RAG/retrieval, APIs, plugins, agents, cloud AI services, vector stores, and the identities that call them – plus governance and shadow AI.
Do you need our model weights?
No. We work at the inference boundary, integrations, identity, and governance layers unless a separate red-team SOW says otherwise.
How is this different from a free self-score?
The calculator is a four-minute Observability Gap self-check. This engagement is analyst-delivered with scored domains and a board-ready roadmap.
Can findings feed managed security?
Yes. Findings map into SENTRY operations and VERITY governance programs when you choose to continue.