HiddenLayer Raises $100M as Enterprise AI Security Race Accelerates

HiddenLayer lands $100M Series B led by M12 and Moore Strategic

HiddenLayer announced a $100 million Series B this morning, bringing total funding to $156 million. Microsoft's M12 and Moore Strategic Ventures co-led the round with participation from existing investors Ten Eleven Ventures and Ridge Ventures. The Austin-based startup builds security tooling that monitors AI models, agents, and the expanding ecosystem of plugins and tools they invoke. CEO Chris Sestito said the capital will fund product expansion across model scanning, runtime protection, and supply chain verification for enterprise AI deployments.

Enterprises are deploying agents faster than security teams can review them

Security teams are playing catch-up as business units ship AI agents into production without standard review cycles. The typical enterprise now runs dozens of agents across customer support, code generation, data analysis, and workflow automation. Each agent pulls in external tools browser automation, API connectors, document parsers creating a supply chain that traditional application security tools cannot see. HiddenLayer's funding reflects how urgent this blind spot has become for CISOs.

The technical gap: agents orchestrate tools that existing scanners miss

An agent is not a static model artifact. At runtime it chains together vector databases, function calling frameworks, and third-party APIs. A prompt injection in a document parser can cascade into unauthorized data access through a downstream API connector. Static model scanning catches poisoned weights but misses runtime tool misuse. HiddenLayer's platform instruments the agent runtime to monitor tool invocations, data flows, and policy violations across the full execution graph.

Competitive landscape crowds with model scanning and runtime startups

HiddenLayer competes with Protect AI, Lakera, Robust Intelligence, and CalypsoAI all raising Series A or B rounds in the last eighteen months. Microsoft, Google, and Amazon have launched native model scanning in their cloud consoles. The differentiation is shifting from model scanning to agent runtime governance. HiddenLayer's bet is that enterprises need a unified policy layer that spans model provenance, tool permissions, and data egress controls rather than point solutions for each layer.

Security implications for multi-agent architectures and tool ecosystems

Multi-agent systems introduce new attack surfaces: agent-to-agent communication, shared memory stores, and delegated tool access. A compromised research agent can instruct a coding agent to exfiltrate secrets through a legitimate GitHub tool. Traditional network segmentation does not apply when agents share context windows. HiddenLayer's approach treats the agent graph as a distributed system requiring identity, least-privilege tool scopes, and audit trails for every tool invocation not just model inputs and outputs.

Build versus buy calculus shifts as agent frameworks standardize

Six months ago, security teams built custom guardrails on top of LangChain or AutoGen. Now frameworks like LangGraph, CrewAI, and OpenAI's Assistants API embed tool calling natively. Custom wrappers break with every framework update. Enterprises are moving toward vendor platforms that integrate at the framework layer rather than the application layer. HiddenLayer's SDKs for Python and TypeScript agent frameworks signal where the market is heading: security as a framework primitive, not an aftermarket wrapper.

HiddenLayer's platform extends beyond model scanning to supply chain verification

The platform now includes a model registry with SBOM generation, a policy engine for tool permissions, and runtime enforcement that blocks unauthorized tool calls. Customers can define policies like "this agent may only call internal APIs" or "no outbound network calls from document processing agents." The enforcement point sits in the agent runtime, not at the network perimeter. This architecture matches how modern agent frameworks actually execute locally or in serverless functions, not behind a WAF.