HiddenLayer built its name scanning model files and detecting attacks on AI at runtime. Cranium AI secures AI and proves it is governed, starting from your code. Here is how they compare across the AI Trust Loop.
HiddenLayer is an AI security platform best known for scanning model files for malware and backdoors, plus runtime detection and response, attack simulation and cloud AI discovery.
Cranium AI is an AI security and governance platform that builds AI‑BOMs from application code, scans model files, enforces policy on live prompts and responses, red teams models, maps AI systems to the EU AI Act, NIST AI RMF and ISO/IEC 42001, and shares evidence through AI Cards. HiddenLayer suits teams focused on model supply chain security; Cranium AI suits teams that must secure AI and prove governance.
Every capability on this page, scored the same way for both platforms and rolled up by AI Trust Loop stage.
Cranium AI17/ 20AI‑BOMs from code, CycloneDX and SPDX
Cloud connectors, per-model AIBOM
Live prompt and response monitoring
Runtime detection, session replay
EU AI Act, NIST AI RMF, ISO 42001
MITRE ATLAS and OWASP mapping
Inline Block, Modify or Pass
Model scanning, Detect by default
AI Cards for vendors and customers
Security dashboard and SARIF exports
Scored across the 20 capabilities compared on this page: Documented = 1, Partial or Preview = 0.5, Unclear or Not documented = 0. Deployment and API rows are not scored. The capability list reflects what buyers ask about AI security and governance, and it includes areas where HiddenLayer leads.
HiddenLayer and Cranium AI overlap, but they were built to answer different first questions. Here is the short version.
| Cranium AI | HiddenLayer | |
|---|---|---|
| Category | AI security and governance platform | AI security platform |
| Built for | CISOs and AI governance leads who need security and proof in one place | CISOs, security teams, AI/ML leaders and developers |
| Company | Emerged from KPMG Studio in 2023. Short Hills, NJ | Founded 2022. Austin, TX |
| How AI is found | Scans code repositories into AI‑BOMs; SIEM-based shadow AI view | Connectors to AWS (Bedrock, SageMaker, AgentCore), Azure ML and Foundry, Databricks |
| Runtime control | Inline Block, Modify or Pass on prompts and responses | Detect, Redact or Block through a self-hosted proxy, API, gateways or agent hooks |
| Security testing | Model-file scanning, Arena red teaming and guardrail testing | AI Attack Simulation and model-file scanning |
| Regulations | EU AI Act, NIST AI RMF, ISO/IEC 42001 | MITRE ATLAS and OWASP in docs; EU AI Act and ISO 42001 not documented |
| Deployment | SaaS with US, EU and AU tenants | SaaS (US, EU), hybrid, self-hosted and disconnected |
Disclosure: Cranium AI publishes this comparison. Every HiddenLayer detail comes from public sources, mainly HiddenLayer’s own website, documentation, marketplace listings and press releases, plus a few third-party publications where cited, checked on October 29, 2026. Where we could not find public evidence, we say “Not documented,” never “No.” Spot something out of date? Tell us and we will fix it.
Organized by the AI Trust Loop: Discover, Observe, Govern, Secure and Prove. Cranium AI is held to the same standard as everyone else: if our docs don’t show it, we don’t claim it.
Find the AI you run
Scans GitHub, GitLab, Bitbucket and Azure DevOps (cloud and self-managed) into AI‑BOMs of models, datasets, packages, agents and static system prompts.1,2
CycloneDX 1.6 AIBOM per scanned model file. SPDX export not documented.7
AgentSensor maps agents, tools, handoffs, guardrails and MCP servers from source code across six agent frameworks.5
Shadow AI view built from connected SIEM activity: AI provider access, AI-enabled SaaS, AI web apps and embedded AI. No endpoint agent or browser extension.10
“Shadow AI” refers to unsanctioned models in connected cloud accounts; no SaaS, browser or SIEM usage view documented.3
Discovery runs from code and SIEM activity. CloudSensor covers Azure security posture; no Bedrock, Vertex or Databricks connectors are documented.11
Watch what it does
Guardian inspects every prompt and response between your AI apps and the model, in Listen or Enforce mode.13,14
Prompts and responses inspected through a proxy, evaluation API, gateway plugins and agent hooks, with session reconstruction.15
Tool Inspector, session reconstruction and agentic analytics: autonomy score, tool invocations and agent-to-agent handoffs.16,17
Agentic Runtime Security replays sessions, screens tool calls and supports threat hunting; docs note it is rolling out by tenant.15
Hold it to policy
EU AI Act, NIST AI RMF and ISO/IEC 42001, with self-attestation and automated answer scoring.18
AI Card reviews with approve, request changes or decline, plus an audit trail. No use-case intake questionnaires.20
Supply chain policy rules, compliance overrides, per-project runtime policies, RBAC and an audit log. No intake or approval workflows documented.21
Profiles set Block, Modify or Pass per signal, separately for prompts and responses, and per agent event type.22
Defend it under attack
Enforce mode runs inline between your app and OpenAI, Anthropic Claude or AWS Bedrock models. New profiles start in Listen.14,24
PII, PHI, PCI, secrets, named entities and a custom blocklist, on prompts and responses.22,26
22 PII entity types including credit card and SSN, plus a generic API key detector. No PHI-specific entities or broad secrets scanner.23
Direct attacks at runtime: jailbreaking, instruction override, prompt leaking, role impersonation, direct command injection and goal hijacking at runtime. Indirect injection is tested in Arena before release.26,27
Proprietary prompt-injection and jailbreak classifier (v5.12, Sept 2026), direct and indirect.28
Deterministic, non-LLM classifiers. Trace shows the text and training samples behind each verdict.29,30
Arena runs fixed and adaptive multi-turn attacks across 9 LLM categories and 26 non-LLM task types, mapped to OWASP.31,27,32
Detects known CVEs in model formats and loaders, referenced in the AIBOM. No application dependency mapping documented.7
Scans imported model files for malicious code and unsafe serialization.
Core strength: 17 formats, malware, code execution, backdoors, tokenizer tampering and container scanning.36
Scans agent config files in repositories (AGENTS.md, CLAUDE.md, .cursor, Copilot instructions) for exfiltration endpoints and hidden Unicode. No endpoint hooks.37
Agent Harness Security hooks into Claude Code, Cursor and GitHub Copilot CLI, deployed through MDM.38
Show the evidence
AI Cards request, respond and publish AI‑BOMs, framework results and vulnerability data; My Vendors tracks every card received.39,40
Third-party AI risk is offered as a professional-services assessment; no product workflow documented.41
Framework and AI System PDF exports, AI Cards and audit trails on review decisions. Runtime evidence is not packaged into AI Cards.42,20
Security dashboard, exportable tables, SARIF and AIBOM outputs, audit log. No board or auditor report packs documented.19
How it runs
“Documented” means the vendor’s public docs or product pages describe the capability as available. “Preview” covers beta, private preview and research preview. “Not documented” means we found no public evidence; it may exist behind a login or on a roadmap. Numbers link to the sources at the foot of this page.
The fastest way to see the difference is to follow one AI system through each platform.
Cloud connectors inventory models, endpoints and agents in AWS, Azure and Databricks, and send each model to the scanner.
Model Scanner inspects files for malware, backdoors and tampering across 17 formats and emits CycloneDX AIBOMs and SARIF.
AI Attack Simulation evaluates system prompts and runs red-team objectives, mapped to MITRE ATLAS and OWASP.
Runtime Security inspects prompts, responses and tool calls through a proxy, API, gateways or agent hooks, and streams outcomes to the SOC.
CodeSensor and AgentSensor scan your repositories and build AI‑BOMs of the models, datasets, packages, agents and MCP servers in your code. A SIEM-fed Shadow AI view and AI Recon cover the AI you didn’t build.
Guardian watches live prompts, responses and tool calls between your apps and the model, with session reconstruction, agent analytics and cost tracking.
Map each AI system to the EU AI Act, NIST AI RMF and ISO/IEC 42001, and turn policy into runtime profiles that decide per signal.
Enforce mode blocks or modifies risky prompts and responses inline. Model files are scanned for malicious code, Arena red teams models before release and Arena Shield tests the guardrails.
AI Cards share AI‑BOMs, framework results and vulnerability data with customers, vendors and auditors, and flag when they fall out of sync.
Buyers trust comparisons that admit what the other platform does well. So do we.
Seventeen documented formats and sixteen detection categories, including graph backdoors and tokenizer tampering, plus container and NIM scanning.36
HiddenLayer is a CVE Numbering Authority and has published widely covered work such as the Policy Puppetry bypass.46
SaaS in US and EU, hybrid, and fully self-hosted or disconnected installs for regulated and federal environments.44
Python, TypeScript and Java SDKs, CI plugins, and runtime outcomes streamed to Splunk, Microsoft Sentinel and webhooks.45,47
Cranium AI maps AI systems to the EU AI Act, NIST AI RMF and ISO/IEC 42001 and shares the evidence through AI Cards. HiddenLayer does not document these frameworks.
Cranium AI scans application code for models, datasets, packages, agents and MCP servers, and exports CycloneDX and SPDX.
Each AI system in Cranium ONE carries its AI‑BOM, framework answers and vulnerability data on one record you can export or share as an AI Card.
AI Cards and My Vendors cover vendor AI as a workflow. HiddenLayer offers third-party risk as a service.
Running both? Both scan model files. HiddenLayer goes deeper on file formats and detection types inside ML pipelines, while Cranium AI adds application-level AI‑BOMs, runtime governance and AI Card evidence. No native integration is documented.
Straight answers, kept current. Last verified .
Ask us anything →An Austin-based AI security company founded in 2022. Its HiddenLayer AI Security Platform has four modules, AI Discovery, AI Supply Chain Security, AI Attack Simulation and AI Runtime Security, plus Agent Harness Security for coding agents.
Malicious code and arbitrary code execution, backdoors including graph payloads, tokenizer tampering, sideloading and known CVEs, across 17 documented formats including Pickle, PyTorch, Safetensors, ONNX, GGUF and Keras. Cranium AI does not scan model files.
Yes, per scanned model file, in CycloneDX 1.6. SPDX export is not documented. Cranium AI builds AI‑BOMs from application code, including agents and MCP servers, and exports CycloneDX and SPDX.
Both are supported. Detect is the default for every rule, and Block or Redact must be configured. Cranium AI’s profiles set Block, Modify or Pass per signal.
HiddenLayer’s documentation maps findings to MITRE ATLAS and OWASP. EU AI Act and ISO 42001 mapping is not documented. Cranium AI maps to the EU AI Act, NIST AI RMF and ISO/IEC 42001.
Yes. HiddenLayer documents self-hosted installs and a disconnected mode in which no customer data leaves the network. Cranium AI runs as SaaS with US, EU and AU regional tenants.
HiddenLayer publishes no price list; enterprise deals are quoted, including through AWS Marketplace private offers.
If you need to secure AI and prove governance in one platform, Cranium AI combines AI‑BOMs from code, inline enforcement, red teaming, framework mapping and AI Card evidence in Cranium ONE.
How we built this page. We read HiddenLayer’s public website, product documentation, developer docs, marketplace listings and press releases, and Cranium AI’s own documentation at docs.cranium.ai, on October 29, 2026. We compare documented capabilities. Announced or preview features are labeled as such, and analyst placements are left out. Product names and trademarks belong to their owners; HiddenLayer has not reviewed or endorsed this page.
See how Cranium AI discovers, observes, governs, secures and proves every model, agent and vendor in your enterprise.
We use essential cookies to run this site, and — only with your permission — analytics and marketing cookies to improve it. You're the hero here; you stay in control. Privacy Policy.
