AI Infrastructure & Governance
The AI stack — models, agent frameworks, vector databases, inference platforms, MLOps tooling — is the youngest market bordering cybersecurity, and the one converging fastest. Security's stake in it runs in both directions, per the two-thread framework of AI Security: the AI stack must be secured (Security for AI), and the AI stack is rebuilding security products (AI for Security). This page covers the adjacency itself — the infrastructure and governance layers as markets, and where their vendor sets collide with security's.
The market and its players
Three layers matter for the convergence. The agent/orchestration layer: LangChain, LlamaIndex, CrewAI and the framework ecosystem — plus Anthropic's MCP as the emerging interoperability standard — define how agents act, which makes them the substrate where agent identity and permissioning problems live (MCP & Agent Identity). LangChain's presence as a founding member of the Agentic SOC Alliance alongside CrowdStrike signals how directly this layer now participates in security's institutions. The observability/evaluation layer: Arize, Weights & Biases (CoreWeave), Galileo, Braintrust, LangSmith — monitoring model behavior, drift and output quality; functionally the observability adjacency (42a) replayed for AI, and the natural home for runtime guardrails. The governance layer: Credo AI, Holistic AI, Fairly AI, and the modules shipping from OneTrust, Vanta and the GRC incumbents (42f) — model inventories, risk assessments, and EU-AI-Act compliance workflows, the regulatory mirror of AI-SPM.
| Layer | Representative players | Where security collides |
|---|---|---|
| Agent frameworks / MCP | LangChain, LlamaIndex, CrewAI | Agent identity, permissioning, MCP security (20b) |
| AI observability / evals | Arize, W&B, Galileo | Runtime guardrails; overlap with AI firewalls |
| AI governance | Credo, Holistic, OneTrust/Vanta modules | The compliance face of AI-SPM (20d) |
| Model/inference infra | Foundation labs, inference clouds | Model theft, weight security, supply chain (20e) |
The convergence mechanics
Every AI-infrastructure layer generates a security market one step behind it — the pattern documented across 2025–26 funding: agent frameworks begat agent-security startups (Zenity $125M, Obsidian $85M at ~$1.1B, the agentic-identity cluster on 03l); model deployment begat AI red-teaming (Mindgard, and the acquired cohort — Protect AI→Palo Alto, Lakera→Check Point, Enkrypt→Anaconda); the npm-scale supply chain begat AI build-chain security (the Sapphire Sleet @mastra compromise on 15). The open question the wiki tracks is who owns each layer's security: the AI-infrastructure vendor adding guardrails natively, the security platform extending in, or the Security-for-AI specialist — three-way contests now visible in AI observability (Arize vs. AI-firewall startups vs. platform modules) and AI governance (Credo vs. OneTrust vs. GRC suites).
The buyer set is still forming, which is the adjacency's defining M&A fact: AI-infrastructure companies with enormous valuations and strategic urgency (the labs themselves, the inference clouds, CoreWeave-style acquirers) have begun buying tooling companies, and nothing prevents them buying security assets next — a fourth acquirer class forming beside platforms, sponsors and adjacents.
Updated 2026-08-16 18:47 UTC · © El Dorado Capital · el-doradocapital.com · Market intelligence for informational purposes only; not investment advice.