This content is currently locked.

Your current Infotech subscription does not include access to this content. Contact your account representative to gain access to Premium SoftwareReviews.

Contact Your Representative
Or Call Us:
+1-888-670-8889 (US/CAN) or
+1-703-340-1171 (International)

Darktrace SECURE AI: Behavioral Anomaly Detection Applied to AI Risk

Technology Note By: Jon Nelson, Infotech

Executive Summary

Darktrace / SECURE AI extends Darktrace’s behavioral detection platform into AI-specific security, applying more than a decade of cross-domain anomaly detection to prompts, agents, and AI activity. It reached general availability in September 2026. It is delivered as SaaS, packaged in Core and Advanced tiers, and sits as one domain on the broader Darktrace platform rather than as a standalone product. Its four capability areas are the discovery of shadow AI, the analysis of prompts and responses, agent and identity monitoring, and the evaluation of AI development risk.

What most distinguishes the product follows from where Darktrace came from. It correlates AI activity with the rest of the estate – email, network, cloud, identity, and OT – in a single incident, so an anomalous agent is seen in context rather than as an isolated alert. It detects sensitive data in a prompt by reading the language itself, without requiring the customer to have labeled that data first. And it judges what is anomalous against a model of the specific organization rather than a global average. These are capabilities a newer entrant cannot easily reproduce, because they rest on mature detection that Darktrace has spent more than a decade building.

The central limitation is that Darktrace’s AI capabilities today are detective, not preventative. The platform sees, correlates, and responds, but at GA that response is identity-based and after the fact: it disables a user once an anomaly has already occurred. It also does not provide an AI Bill of Materials, model supply-chain scanning, adversarial testing, or compliance mapping beyond the EU AI Act – capabilities that sit outside what a behavioral engine does. Prevention and deeper governance are described as roadmap directions but are not available at GA.

Darktrace is best suited to three buyers. The first needs to get control of AI risk quickly but has not done the multiyear work of data classification. The second is an existing Darktrace customer extending into AI. The third has made fast, correlated detection across a broad estate its priority. It is a weaker fit, or requires augmentation, for buyers who need to prevent AI actions inline, for buyers whose priority is governance, compliance, and model integrity, and for buyers who need to map where sensitive data lives at rest. The decision turns less on whether Darktrace is capable and more on whether the buyer’s most urgent need is to see AI risk across the estate now or to prevent specific actions and satisfy governance obligations today.

Introduction

No CISO is choosing how AI enters their organization. Leadership mandates the transformation, the C-suite announces that the company is adopting one assistant or another, and the deployment happens everywhere at once. Engineering stands up agents and MCP servers. A business unit buys a copilot. Someone connects a frontier model to a production dataset. Some of it is sanctioned. Most of it is not. By the time the security team assembles a picture of what is running, that picture is already out of date. Shadow AI is the default state, not an edge case, and the security team is told to go secure it after the fact.

The market has answered that pressure with noise. Every data security vendor, every network vendor, and a wave of recent startups now sells an AI security platform, and each one arrives from the corner of the market it already occupied. Vendors from data security lead with classification and data-loss enforcement. Vendors from the network lead with traffic and posture. Startups lead with the single capability they were built around. For the buyer, this creates a second problem on top of the first. AI is already everywhere, and the tools to secure it all describe the same category in the language of wherever they came from, so it is not obvious which framing matches the risk that matters most.

What this produces is a buyer under three simultaneous pressures. AI is already in the environment, usually without governance or signoff. Accountability for securing it still rests with the CISO. And the security team’s understanding of how these systems work rarely matches what it is being asked to govern. No platform fully resolves that. The best a platform can do is give a stretched security team the visibility, detection, and response to become effective on an AI footprint that is already sprawling around them.

Darktrace enters this space from a different corner: behavior. It extends a behavioral detection engine it has run across email, network, cloud, identity, and OT for over a decade. The same anomaly-based approach it has long applied to the rest of the estate now applies to prompts, agents, and AI activity. What that approach delivers, where it differentiates, and which buyers should evaluate it is the subject of this note.

Product Overview

Darktrace / SECURE AI is delivered as a SaaS offering, accessed through the Darktrace Platform Portal, and reached general availability in September 2026. It can be purchased and deployed on its own, but it is not an isolated product: it is one domain within the broader Darktrace platform, sitting alongside the vendor’s People and Infrastructure coverage on a shared behavioral engine and a common data layer. The AI-specific controls draw on the same anomaly detection, natural-language analysis, and autonomous response infrastructure already built into the platform. How much of that platform a buyer purchases determines what SECURE AI delivers. Its native prompt analysis and sensitive-data detection work from the product’s own integrations, so a standalone buyer gets them in full. However, its cross-domain correlation depends on other Darktrace domains being present to correlate against.

The offering is packaged in two tiers, Core and Advanced, with a third tier, Complete, described as a future release. Core is a full SaaS package that installs in minutes and works entirely through third-party integrations: connections to frontier model providers, MCP, and SaaS applications, drawing on prompt data and AI-related event data from those sources. Advanced deploys Darktrace’s own sensors across endpoint, network, cloud, and OT, which the vendor positions as enabling deeper shadow AI coverage than integrations alone can provide. Core is designed for rapid trials and marketplace distribution; Advanced assumes the buyer is willing to deploy sensor infrastructure to widen coverage.

The platform is built around four core capabilities: the discovery of shadow AI, the analysis of prompts and responses, agent and identity monitoring, and the evaluation of AI development risk. Data reaches the platform through two paths. Prompt and response data from sanctioned applications is collected through provider APIs, including the OpenAI and Anthropic compliance APIs and Microsoft change logs for Copilot. Shadow AI discovery draws on network, endpoint, and email signals, with integrations to secure-access platforms such as Zscaler and Microsoft GSA available at GA.

Two capabilities are worth stating up front because they frame how the product should be evaluated. Sensitive-data classification is performed natively, using the natural-language analysis developed for the vendor’s email security product, so it does not depend on the customer having labeled its data. Response actions, by contrast, are narrow at GA: they are identity-based, executed through Microsoft 365, and limited to logging out or disabling a user. Broader telemetry sources and enforcement hooks, including OpenTelemetry ingestion and inference-level response for frontier model providers, are on the roadmap but not available at GA.

Features and Capabilities

Core Features

The table below measures Darktrace / SECURE AI against the standard capability set for the AI security and governance category, at the pillar level. Coverage is stated as of the September 2026 GA release. Capabilities the vendor described as roadmap are flagged as such, and figures or claims the vendor provided that were not independently validated are marked vendor-reported.

Capability

Coverage

AI Asset Discovery and Inventory

Discovers sanctioned and unsanctioned AI tools, agents, and identities, drawing on network, endpoint, and email signals rather than a single API connection. Discovered services appear in an inventory with detection counts, entities involved, last-seen time, and an approved or unapproved status. The vendor states passive network detection can detect AI use on unmanaged and bring-your-own devices where the traffic crosses a monitored network (vendor-reported). Broader discovery sources, including a secure browser extension, endpoint AI discovery, and AI traffic fingerprinting, are on the roadmap.

AI Bill of Materials (AIBOM)

Not available at GA. The product does not produce an inventory of AI components, models, libraries, or dependencies, and AIBOM was not described as on the roadmap.

AI Security Posture Management

Partial at GA. The product detects misconfiguration and over-permissioning in AI development activity. The vendor stated that broader posture management, extending beyond AI to the wider environment, resides in its separate cloud security product. Agent misconfiguration detection is on the roadmap.

AI Supply Chain Risk

Not available at GA. The product does not scan code libraries or model repositories for CVEs or known vulnerable components. It can detect AI-related supply chain activity behaviorally, such as an application making unexpected outbound calls to an AI service, but this is anomaly detection, not component scanning.

Adversarial Testing

Not available at GA. The product does not perform AI red teaming or adversarial testing of agents, models, or MCP servers, natively or through a named partner, and this was not described as on the roadmap.

AI Gateway and Runtime Guardrails

Limited at GA. The product does not operate as an inline gateway brokering prompts or agent actions. It analyzes prompts and responses collected through provider APIs and can respond after a detection fires. Inference-level response hooks for frontier model providers are on the roadmap.

MCP Coverage

Partial at GA. MCP activity is ingested as a data source and can appear in shadow AI discovery. Dedicated MCP server discovery and inventory is on the roadmap. The product does not perform adversarial testing against MCP servers.

Compliance and Policy Enforcement

Partial at GA. Customers can upload an acceptable-use policy document, which the product parses into discrete rules and measures the live environment against, reporting an alignment score and per-rule violation counts. Assessment against the EU AI Act is available through a legislation manager. Mapping to other frameworks, including the NIST AI Risk Management Framework and ISO/IEC 42001, was not available at the time of the briefing. Policy handling measures and alerts on violations; it does not block them inline.

Third-Party AI Risk Management

Not available at GA. The product does not provide vendor onboarding workflows, third-party AIBOM collection, or vendor questionnaire analysis.

AI Activity Monitoring and Lineage

Available at GA. The product captures AI activity across prompts, responses, sessions, agents, and identities, and applies anomaly detection across them. Its distinguishing strength is correlating AI activity with signals from the wider estate (email, network, cloud, identity, OT) in a single incident. Agent permission and access visualization is present; full user-to-agent-to-model-to-data lineage of the kind offered by data-context platforms is not the product’s basis.

Identity and Agent Governance

Partial at GA. The product inventories AI agents, the identities and credentials they use, and the permissions attached to them, and visualizes what an agent can reach. Response actions are identity-based, executed through Microsoft 365, and limited to logging out or disabling a user. Deeper governance, including permission mapping, blast-radius analysis, and agent pattern-of-life baselining, is on the roadmap.

The gaps are not random. The capabilities Darktrace covers well are the behavioral ones: discovery from signals across the estate, activity monitoring, anomaly detection, and correlation. The capabilities it does not cover are the ones that sit outside behavior: bill-of-materials, supply chain scanning, adversarial testing, and broad compliance mapping. This is a coherent product with a clear center of gravity, not a platform with scattered gaps.

A reader should be careful not to simply chase a long feature list. A vendor may check the red-teaming box with testing too shallow to catch a real-world prompt injection. The more useful question is one of sequence. Posture management matters, but a posture finding on an agent you did not know existed is noise; you cannot manage the posture of what you have not yet discovered. Discovery comes first, and governance depth builds on top of it. How well Darktrace’s coverage and its roadmap fit that logic is taken up in the Analyst Perspective and Fit Guidance that follow.

Differentiating Features

One behavioral engine across the enterprise. Many tools entering this category analyze AI in isolation: they see the prompt, the agent, or the model, and nothing around it. Darktrace’s behavioral heritage provides a much more complete view: email, network, cloud, identity, and OT. An AI anomaly is not an isolated AI alert. The correlation is already done, so a SOC analyst is not pivoting across separate tools to reconstruct what happened, which cuts alert fatigue and shortens response time.

Consider a set of compromised credentials driving an AI agent. A tool watching only the AI layer sees an agent issuing unusual prompts and flags it. Darktrace sees the same prompts alongside the login behavior that does not match the user, the network connections the session opened, and the data it reached, and reads them as one incident. What a single-layer view can only flag as ambiguous, the behavioral engine resolves into a clear compromise.

Sensitive-data detection without prior labeling. Most approaches to catching sensitive data in AI depend on that data having been classified or labeled first, and in practice most organizations have not done that work. PII is the easy case, recognizable by pattern almost anywhere. The harder case is proprietary data that carries no standard signature: a construction firm’s bid, a health system’s trial data, an M&A document. It is sensitive because of what it is to the business, not because it matches a known format, and it is rarely labeled.

Darktrace classifies this content by reading the prompt itself, using the natural-language analysis built for its email security product. Where labels already exist like Microsoft Purview, it can use them. But it does not depend on them. A customer can write a rule for the data they care about and have it detected in a prompt without tagging that data first. It does not map where sensitive data lives at rest, which is what a classification-led platform is built for. For the specific problem of sensitive content leaving in a conversation with an AI, label-free detection is the more practical starting point.

A model of your environment, not everyone’s. When most vendors say they have an AI model, they mean one model, trained on pooled data and shipped to every customer. That model defines normal as an aggregate across its whole training population. Darktrace takes the opposite approach: the engine learns each business from its own data, with no pre-training, so normal is defined for that specific organization. What it flags is what is unusual and significant for you, not what deviates from a global baseline.

The difference is what each model misses. A remote administration tool that is legitimate at some companies reads as normal to a pooled model, so a new one appearing in your environment is waved through, even when it is an attacker’s foothold. A model trained only on your environment sees a tool that has never been here before and flags it. The tradeoff is that this model must learn the environment before it can judge it, so value builds over the first weeks rather than arriving complete on day one.

Analyst Perspective

Every vendor in this category enters from their particular area of strength. Data security vendors extended classification and enforcement into AI. Startups built around a single capability. Darktrace arrives from behavioral analytics: more than a decade of learning what is normal for an organization and flagging what is not, across email, network, cloud, identity, and OT. The buyer’s task is not to find the platform with the most features. It is to work out which entrant addresses their specific risks.

What Darktrace does, it does quickly: it stops the bleeding. It helps a security team get its arms around the things that matter most right now, the AI in the environment they do not know about, the sensitive data going out the door in a prompt, and the agentic AI going off the rails. It does this without a data-classification program in place, reading sensitive content from the language of the prompt rather than from labels the organization has probably not applied. For a buyer who needs exactly this, who needs it now, and who has not done the years of groundwork that a classification-led platform assumes, Darktrace is a compelling choice.

What it does not do is just as important. It does not prevent data leakage or incidents; it detects them, and its response at GA is after the fact, disabling a user once the event has already happened. For a buyer in a regulated or high-consequence environment who needs to stop an action before it happens, a detective control does not meet the need.

It does not tell you where your sensitive data lives across the estate. It observes that data in motion through a prompt, not at rest.

That is the trade the buyer must make. Darktrace is strong at seeing and correlating AI risk across the whole estate, and it delivers that value fast, on the environment as it actually is. It is not, today, built to prevent. But detection is the right place to start – you cannot stop what you cannot yet see – and a buyer who adopts Darktrace to get its arms around AI risk now is not buying a dead end. The platform’s stated direction is toward more preventative control, including intervention at the point of inference, so the buyer who needs visibility today can reasonably expect the enforcement side to mature over time.

Conclusions and Fit Guidance

Strong Fit

Buyers who need to get control of AI risk quickly and have not done data classification. This is the clearest fit. For an organization where AI is already in use, sensitive data is already moving into prompts, and the multiyear work of discovering and labeling data has not been done, Darktrace delivers value without that groundwork. It reads sensitive content from the language of the prompt rather than from labels, and it detects shadow AI, prompt-level data exposure, and anomalous agent behavior. For the buyer who needs to stop the bleeding now, this is a strong starting point.

Existing Darktrace customers extending into AI. The cross-domain correlation that most distinguishes the product depends on the rest of the estate being visible to the same engine. A buyer already running Darktrace across email, network, cloud, identity, or OT realizes that advantage immediately, and the Advanced tier’s deeper shadow AI coverage assumes those sensors are already deployed. For this buyer, AI becomes one more domain on a platform already in place rather than a new tool to stand up.

Buyers whose priority is detection and rapid response across a broad, multidomain estate. Where the need is to see and correlate AI risk alongside the rest of the environment, catching the anomalous agent, the risky prompt, and the unsanctioned tool as parts of a single picture, the behavioral engine is on its home ground. Buyers who value fast, correlated detection over inline enforcement will find the product well matched to how they work.

No-Fit or Requires Augmentation

Buyers who require inline prevention. For organizations in regulated or high-consequence environments where the requirement is to stop a sensitive prompt or an unauthorized agent action before it completes, Darktrace’s GA capabilities do not meet the need. Its response is detective and after the fact, disabling the user once the event has already occurred. Prevention is described as a roadmap direction, but a buyer with this requirement today should evaluate the product on what it does now, not on what it may do later.

Buyers whose priority is governance, compliance, and model integrity. Darktrace does not provide an AI Bill of Materials, model supply-chain scanning, or adversarial testing at GA, and its compliance mapping is limited to the EU AI Act, with other frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 not available at the time of the briefing. A buyer whose critical need is regulatory framework mapping, audit evidence, model provenance, or supply-chain assurance should look to a platform built for that work or plan to augment.

Buyers who need to map where sensitive data lives at rest. Darktrace detects sensitive data in motion through a prompt. It does not discover and map sensitive data across the estate the way a classification-led platform does. A buyer whose primary problem is knowing where sensitive data resides, rather than catching it as it moves into an AI tool, is buying the wrong shape of product.

The right question is not whether Darktrace is the most complete AI security platform. In a category this young that judgment could change next week. What decides the fit is whether the buyer’s most urgent need is to see and correlate AI risk across the estate today or to prevent specific AI actions and satisfy governance and compliance obligations. For the first, Darktrace is a strong and fast-moving choice. For the second, it is either the wrong fit today or a starting point that will require further augmentation.

Latest Technology Notes

All Technology Notes
Over 100 analysts waiting to take your call right now: +1 (703) 340 1171