ThreatIntelligence-IncidentResponse

MITRE ATLAS in Elastic Security: AI detection coverage


Elastic Security 9.6 supports MITRE Adversarial Threat Landscape for Artificial-Intelligence Systems (ATLAS), MITRE’s knowledge base for attacks on AI systems, in addition to MITRE ATT&CK. ATLAS covers prompt injection, jailbreaks, stolen model API keys, and poisoned agent tools. In Elastic Security 9.6, prebuilt detection rules for AI platforms and generative AI (GenAI) tooling will ship with ATLAS mappings already applied, so rules provide the AI-related context out of the box. MITRE recently announced that ATT&CK Enterprise will add AI-enabled systems as a new platform, drawing in part on ATLAS content. However, ATLAS remains MITRE’s dedicated, frequently updated AI threat knowledge base. A new coverage table shows which ATLAS tactics and techniques your rules already address, scored in its own view rather than averaged into your ATT&CK numbers. The rule creation form and Elastic Agent Builder also understand ATLAS. Behind it all is a new managed MITRE data service in Kibana.

We’ve supported MITRE ATT&CK since the initial version of Elastic Security, and we’ve extended that support over the years with ATT&CK mappings on prebuilt rules and features, like the coverage overview page. 

What is MITRE ATLAS?

MITRE ATLAS is a knowledge base of the tactics and techniques that adversaries use against AI-enabled systems. It’s maintained by MITRE and shaped by real-world observations and security research from across the industry and academia. ATLAS is modeled directly on ATT&CK, so the structure will feel familiar to users: tactics to describe what an adversary is trying to achieve, plus techniques and sub-techniques to describe how they go about it.

MITRE ATLAS versus MITRE ATT&CK

ATLAS and ATT&CK use the same structure, but they describe attacks against different kinds of systems.

MITRE ATLAS

MITRE ATT&CK

What it covers

Models, training data, and agents

Hosts, networks, identities, and cloud infrastructure

Structure

Same structure, reusing ATT&CK tactics and adding AI-specific ones, like AI Attack Adaptation and AI Model Access

Tactics, techniques, and sub-techniques

Update cadence

Monthly content releases

Roughly twice a year

In Elastic Security

Prebuilt rule mappings, coverage overview tab, rule form, and Agent Builder

Prebuilt rule mappings, coverage overview tab, rule form, and Agent Builder

The goal of ATLAS is to give defenders the same thing that ATT&CK did a decade ago: a common, evidence-based vocabulary for a class of threats that they can use to assess their own vulnerabilities. And ATLAS complements ATT&CK, rather than replacing it. Attacks on AI systems rarely happen in isolation. An adversary might use familiar intrusion techniques to get access and then, once inside, turn to AI-specific techniques. Mapping detections to both frameworks, like some of the detection rules we’re shipping, gives users a more complete picture of an attack, including the infrastructure it propagates and the AI systems it targets.

For Elastic Security users, this means a clear view of how well your detections cover the AI systems in your environment, measured on its own terms instead of mixed into your ATT&CK numbers. As your organization adopts large language models (LLMs) and AI services, in addition to autonomous agents, you can see which AI-specific threats your rules already address and find gaps before an attacker does. Our prebuilt detection rules for AI platforms and GenAI tooling now carry ATLAS mappings alongside ATT&CK, so this coverage view is populated from day one.

How MITRE ATLAS works in Elastic Security

In Kibana’s 9.6 release, we will implement support for ATLAS in the ways that customers have come to expect from the MITRE ATT&CK data. A single rule can carry mappings from both frameworks in its threat field. Plus, they sit side by side in the UI, so a user can view coverage counts or external links to MITRE reference data in the same way that they always have. Beyond the addition to our prebuilt rule mappings, we’ve also introduced: 

    A new coverage overview table. Mirroring MITRE’s own matrix visualization, so customers can see which threat tactics and techniques their rules target.

  • Manual rule create/edit UI support. The classic detection rule creation form, updated to give customers a visual way to add both ATT&CK and ATLAS mappings to their custom rules or to edit existing prebuilt rules to enrich potential alerts.

  • AI rule creation. The newly introduced rule creation workflow now supports adding both ATT&CK and ATLAS mappings to rules, where appropriate.

  • Agent Builder workflows. Elastic Security’s Agent Builder skills now understand ATLAS, too, so you can ask an agent which AI threats your installed rules cover or have it recommend prebuilt rules for specific ATLAS tactics and techniques.

How Kibana serves MITRE ATT&CK and ATLAS data as Saved Objects

In the past, we’ve stored our own version of the MITRE ATT&CK framework within Kibana, derived from artifacts released by MITRE itself via Structured Threat Information Expression (STIX) datasets. We’d generate a local artifact via a script and ship it within our stack or serverless release bundles. This method largely worked for users, but it also limited us in some key ways, in that we couldn’t easily support specific ATT&CK versions or more than one MITRE framework. 

The generated MITRE artifacts were large enough that we had to lazily load them in the UI so they didn’t blow up our webpack chunk size. If we wanted to search through the entities at all, we’d have to load in the entirety of the data and search using basic array operations. The updating process of these artifacts was also tied to stack releases, causing occasional delays and version misalignments for customers who pick up new prebuilt rules as soon as they ship.

With MITRE ATLAS support, we saw a way to both add a new feature and tackle some of this long-standing tech debt by introducing a new managed MITRE data system. We still generate a similar artifact from the STIX datasets, but we now ingest the entities into Saved Objects to serve them to any consuming features via a data client and an API. This allows us to query and filter MITRE data in the same way that we do any other Saved Object, supporting a number of new feature opportunities, some of which (like MITRE ATLAS) have already been implemented. 

We’ve also expanded the fields on the entity schema itself, adding new information, like an entity’s full description, its revocation status, its version, and more. This provides us a much richer context to work with when building tools for users to integrate MITRE data in their workflows and to keep it up to date in their custom rule mappings.

The managed approach also opens the door to move away from packaging the artifact with Kibana releases themselves and to follow the more frequent prebuilt rule release lifecycle and delivery method, so the incongruencies of version mismatches won’t be an issue for those customers who want the latest and greatest as soon as they’re released. This matters more for ATLAS, which ships monthly content releases.

What’s next for MITRE framework support in Elastic Security

Our goal is to keep building on the managed MITRE data service, giving customers the most up-to-date threat information in the workflows that they already use and building new ways to understand their coverage. A few of the areas we’re exploring include:

    More frameworks. The data model was built to hold multiple MITRE matrices, which makes it straightforward to bring in others, like ATT&CK for Industrial Control Systems (ICS), as customer needs grow.

  • Smarter search for AI-assisted workflows. Every MITRE entity now carries its full description, providing valuable context for agent reasoning. We’re experimenting with semantic search, so AI-driven features, like AlertZero, can use the latest tactics and techniques most relevant to a given behavior, instead of relying on a model’s possibly outdated knowledge base. This update provides higher-quality output on a broader set of models.

  • Faster framework updates. Decoupling MITRE data from Kibana releases so new ATT&CK and ATLAS versions arrive on the same schedule as the prebuilt rules that reference them. This matters as ATLAS ships monthly and ATT&CK’s AI platform arrives in 2027.

We’d love to hear how you’re putting ATLAS to work. If you have feedback on the new coverage view, the rule form, or the mappings in our prebuilt rules, reach out via our Slack Community or the Discuss forum.

Try MITRE ATLAS coverage in Elastic Security

To experience the full benefits of what Elastic has to offer for detection engineers, start your Elastic Security free trial. To learn more, visit Elastic Security.

The release and timing of any features or functionality described in this post remain at Elastic’s sole discretion. Any features or functionality not currently available may not be delivered on time or at all.



Source link