DeepInspect vs HiddenLayer: Runtime Enforcement and Model Scanning Compared for Enterprise AI Programs

DeepInspect and HiddenLayer occupy adjacent positions in the enterprise AI security market. The two products started from different angles: DeepInspect from runtime HTTP enforcement, HiddenLayer from adversarial ML research and model scanning. The shortlist conversation usually centers on the same question: which placement produces the per-decision audit record the program needs.

TL;DR

DeepInspect is an identity-aware HTTP-proxy that authenticates the caller against the corporate IdP and commits per-decision records on every request. HiddenLayer's AIDR (AI Detection & Response) covers runtime telemetry and adversarial-ML detection, and the Model Scanner covers artifacts before they enter production. Pick DeepInspect if the program centers on identity-bound enforcement records under Article 12 or HIPAA audit obligations. Pick HiddenLayer if adversarial ML detection and model artifact scanning are the primary controls. Programs often run both.

Where DeepInspect sits

DeepInspect sits inline on the HTTP path between authenticated users or agents and any LLM. The proxy terminates TLS at the inspection layer, authenticates against the corporate IdP, classifies the prompt content, evaluates policy against identity and classification, and commits a per-decision audit record before the model receives the request. The records carry identity, classification, policy version, decision, timestamp, and an integrity signature on a tamper-evident series.

The product is purpose-built for the runtime enforcement boundary. It does not scan model artifacts before they enter production. The artifact-side coverage belongs upstream.

HiddenLayer: where it sits

HiddenLayer ships two named surfaces. The Model Scanner inspects model artifacts at the registry or pipeline boundary, catching malicious payloads in Pickle files, malicious LoRA adapters, and known-vulnerable dependency patterns. AI Detection & Response (AIDR) monitors runtime AI traffic and inference telemetry, looking for adversarial ML patterns including model evasion, model extraction, membership inference, and prompt injection variants.

HiddenLayer's adversarial-ML lineage shows up in the AIDR detection catalog. The platform's strength is the breadth of adversarial-ML detections, including model-extraction patterns the prompt-injection-focused tools do not natively catch. The audit record on the AIDR side is shaped around the detection event, with prompt content and policy state attached when the integration is wired through.

Feature comparison

| Axis | DeepInspect | HiddenLayer | |---|---|---| | Primary surface | Runtime HTTP enforcement | Runtime AIDR + Model Scanner artifact side | | Primary placement | HTTP proxy at request boundary | SDK / telemetry collectors plus artifact scanner | | IdP integration | Built in at proxy | Application or platform-side integration | | Identity binding on every record | Yes by default | When the integration carries it through | | Classification | Deterministic categories (PII, PHI, source code, customer, custom) | Adversarial-ML detection catalog plus content classification | | Adversarial ML coverage | Pattern detection for prompt injection plus policy on tool calls | Broad catalog including model extraction, membership inference, evasion | | Multi-provider coverage | Yes | Yes | | Tamper-evident record series | Yes (signed) | Available as platform feature | | Article 19 natural-person field | Yes by default | Yes when wired through | | Model artifact scanning | No | Yes via Model Scanner | | Latency overhead | Under 50 ms in internal testing | Detection-side, comparable on proxy path |

Pick HiddenLayer if

• The program's primary risk model centers on adversarial ML against production models, including model extraction, evasion attacks, or membership inference.
• The artifact pipeline runs many self-hosted open-weight models and the scanner needs first-class coverage.
• The detection-and-response posture (alert, triage, response playbook) matters more than per-request enforcement at the boundary.
• The team is comfortable owning the integration that carries identity from the application to the detection telemetry.

Pick DeepInspect if

• The program centers on identity-bound per-request records under EU AI Act Article 12, Article 19, or HIPAA Security Rule audit obligations.
• The deployment spans multiple LLM providers on the same policy surface and the record series needs to be canonical.
• Enforcement happens at the request boundary before the model receives the request, not after the model returns the response.
• The IdP integration belongs at the inspection boundary rather than inside each application team's code.

Combining the two

A program with both adversarial-ML risk and audit record obligations runs the two products on different surfaces. HiddenLayer's Model Scanner inspects artifacts at the registry boundary. DeepInspect enforces policy at the runtime HTTP boundary and carries the canonical record series. HiddenLayer's AIDR can supply additional adversarial-ML detections that feed into the same SIEM or SOAR pipeline the DeepInspect records flow into. The two products report into the same audit program through different evidence pipelines.

Regulatory framing

EU AI Act Article 12 requires automatic recording of events over the lifetime of the system. Article 19 specifies identification of natural persons involved. Article 99 sets penalties at €15 million or 3% of global annual turnover for high-risk non-compliance. The August 2, 2026 deadline applies to high-risk AI systems including credit scoring, employment screening, education access, and biometric identification.

The auditor samples decisions from the runtime record series and asks who made the request, what classification the prompt carried, what policy applied, and what decision the system returned. The HTTP-proxy placement with IdP integration produces those fields on every record. The adversarial-ML detection catalog contributes evidence on a different question (which model behaviors looked anomalous), which is parallel to the audit record series.

Pricing approach

Both vendors quote against the specific deployment after scoping. DeepInspect prices per protected endpoint and request volume tier. HiddenLayer prices across the Model Scanner and AIDR modules with platform tiers. Public price lists are not available for either product.

DeepInspect

DeepInspect is the runtime HTTP-proxy enforcement gateway. The proxy authenticates the caller against the corporate IdP, classifies the prompt content, evaluates policy against identity and classification, and commits a per-decision audit record before the response returns to the application. The records carry the fields EU AI Act Article 12 and Article 19 expect on the series HIPAA Security Rule references.

For programs comparing DeepInspect to HiddenLayer, the framing I find useful is to separate the artifact question from the runtime enforcement question and pick the placement that produces the record series the audit program references. The proxy placement is the one that produces identity-bound records across multiple LLM providers on a shared series.

If you are facing the August deadline, let's talk.