Overview

HiddenLayer sells an AI security platform organised around four modules: AI Discovery for locating shadow AI across environments, AI Supply Chain Security for validating model integrity before deployment, AI Attack Simulation for continuous adversarial testing and AI Runtime Security for detection and response in production. The architecture is agentless and model-agnostic, and the company markets it on the basis that it never needs access to customer weights or training data, which is the point that gets it into finance, healthcare and federal accounts.

The supply chain piece is the one worth an engineering lead's attention. Model files are executable in practice. A pickle-serialised checkpoint pulled from a public hub can run arbitrary code on load, and HiddenLayer's own ShadowLogic research showed backdoors can be embedded in a model's computational graph in a way that persists through format conversion and fine-tuning. Their scanner targets exactly that class of problem, checking artefacts for malware, exploits and integrity before they enter a registry. Partnerships with AWS and Databricks put it near where models actually land.

On maturity, this is a real security company rather than a wrapper, founded in 2022 in Austin with a $50M Series A behind it and a research team that publishes work other vendors then have to respond to: Policy Puppetry, TokenBreak, EchoGram, ShadowGenes. But the commercial model is unambiguously enterprise. Nothing is self-serve, no price is published and evaluation means a demo booking. If the question is whether your team can start scanning models this week, the open source ModelScan or picklescan will get you further faster, and HiddenLayer becomes relevant when the problem is organisational rather than technical.

Key Features

✓ Model Scanner for malicious code injection, pickle deserialisation payloads and architectural backdoors in model artefacts before they reach a registry

✓ AI Discovery for finding shadow AI assets across environments

✓ AI Attack Simulation for continuous adversarial testing against deployed models

✓ AI Detection and Response for runtime monitoring without needing access to weights or training data

✓ Agentless, model-agnostic deployment that does not require exposing proprietary IP

✓ Coverage extending to agentic and MCP surfaces, plus agent harness security

My Take

The research output is the reason to take this one seriously. Policy Puppetry, the technique that reformats a jailbreak to look like an XML or INI policy file and bypassed guardrails across every major frontier model, came out of this team. So did ShadowLogic, a backdoor planted directly in a model's computational graph that survives format conversion and downstream fine-tuning. The product is four modules: discovery, supply chain, attack simulation and runtime. Scanning model artefacts for pickle payloads and architectural backdoors before they enter your registry is the piece with the clearest value, because a pulled Hugging Face checkpoint is arbitrary code execution waiting to happen. There is no self-serve tier and no published price, so you cannot try it on a Friday afternoon. That alone rules it out for smaller teams.