ANCHOR: Schema-Agnostic Threat Knowledge Graph Construction via Hybrid Ontology Discovery

arXiv:2606.01208 · cs.CR · Submitted 2026-05-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: I'm Nadia.

Elias: And I'm Elias, here with Priya.

Nadia: We're looking at the paper: "ANCHOR: Schema-Agnostic Threat Knowledge Graph Construction via Hybrid Ontology Discovery".

Elias: Here's the gist of the paper. Cyber threat intelligence (CTI) reports now serve as essential resources for capturing adversary tactics, techniques, and procedures observed in modern attack campaigns While traditional CTI platforms reduce this intelligence to isolated indicators through fixed schemas such as STIX, ontology-based representations preserve the semantic relationships needed for structured threat analysis In this paper, we present ANCHOR, a schema-agnostic CTI knowledge graph construction system that bridges LLMs and formal ontology schemas Experimental results on the UCO, STIX, and MALOnt schemas show that ANCHOR outperforms existing baselines in ontology typing and schema compliance In addition, ANCHOR with a local LLM closely matches enterprise LLM typing performance, enabling privacy-preserving CTI analysis with high fidelity Pruning redundant reasoning tokens The paper introduces ANCHOR as a schema-agnostic CTI knowledge graph construction system that addresses three key challenges in existing approaches: (i) schemaspecific pipelines that require manual reconfiguration whenever the schema changes, (ii) prompt-based schema inclusion that fails to scale on large ontologies such as UCO, and (iii) reliance on enterprise LLM APIs that conflicts with privacy constraints when integrating sensitive internal incident data.

Nadia: The core of ANCHOR is hybrid ontology discovery, a search-and-navigate mechanism that dynamically explores largescale ontology schemas, combined with SHACL-based validation to enforce schema-compliant type assignments.

Elias: How the training signal is built ANCHOR utilizes a hybrid ontology discovery mechanism that combines embedding-based semantic search with LLM-guided recursive navigation to align extracted elements to formal ontology classes or properties.

Nadia: The entity ontology typing phase employs Algorithm 1, which first performs an embedding-based search and then, if the confidence falls below a predefined threshold τentity, it activates hierarchical recursive navigation over root classes Uroots.

Paper summary: Nadia: So, looking at the whole paper, "ANCHOR: Schema-Agnostic Threat Knowledge Graph Construction via Hybrid Ontology Discovery," it seems the main achievement is decoupling extraction logic from schema definitions. They built a system that can handle different OWL and SHACL ontologies without needing a manual pipeline reconfiguration every time.

Elias: And they tackled the scaling problem on large ontologies by dynamically discovering only task-relevant ontology fragments instead of putting the whole schema into the language model prompt, which lets you keep each query within a manageable context window.

Priya: I think what this means for threat intelligence is that we can finally build a structured knowledge graph where we capture those subtle semantic relationships between attack tactics and techniques, even when our internal data uses different organizational vocabularies.

Nadia: Exactly. They showed that hybrid ontology discovery combined with SHACL-based validation creates a system that preserves semantic connectivity while maintaining high typing fidelity across different threat intelligence schemas.

Elias: And they demonstrated that by offloading structural reasoning from the LLM to the schema, they reduced the performance gap between local language models and enterprise backbones. That’s a big point for privacy.

Priya: It validates the use of local LLMs for this kind of sensitive work because it shows you can get high fidelity results without exposing proprietary incident data to external APIs.

Nadia: So, "ANCHOR: Schema-Agnostic Threat Knowledge Graph Construction via Hybrid Ontology Discovery" is about creating a flexible, private way to build structured threat intelligence knowledge graphs that works across different standards.

Conclusion: Nadia: So, we've seen how ANCHOR builds this knowledge graph using hybrid discovery and SHACL validation across different threat schemas. Now we're at the end of the paper talking about what this actually means for us.

Elias: Yeah, so looking at that title, "ANCHOR: Schema-Agnostic Threat Knowledge Graph Construction via Hybrid Ontology Discovery," it really hammers home that this system isn't locked into one specific way of organizing threat data.

Nadia: Exactly. It’s not just about fitting data into a fixed box anymore; it’s about building something flexible that can handle different vocabularies and structures for incident response.

Elias: The authors show they tackled the problem of schema dependency head-on, which is huge because most CTI tools get stuck needing manual work every time you switch from one standard to another.

Nadia: And they did that by using this hybrid ontology discovery method—combining embedding search with recursive navigation—to find the right structure dynamically. It’s not just looking up keywords anymore; it's exploring the whole schema space.

Elias: They also addressed the privacy issue directly, showing that you can still get high performance even when you use a local language model instead of a massive enterprise one. That’s a practical win for handling sensitive internal data.

Nadia: It means we can finally create threat knowledge graphs that are both semantically rich and compliant with different organizational standards without relying on those big, centralized enterprise systems for every single query.

Elias: The core idea is that schema-grounded reasoning is more important than just having a super smart language model; the structure itself guides the intelligence extraction.

Nadia: It confirms that for deep analysis, you need to respect the formal definitions of how things are related in cybersecurity, not just rely on pattern matching.

Elias: It’s promising because it shows a path toward making CTI analysis more adaptable and keeping sensitive operational data secure while still getting high-quality structured intelligence out of it.

Nadia: The question now is how quickly we can get this kind of flexible, private graph construction into our real-world threat hunting pipelines.

Ministry of National Defense Incheon International Airport Corporation Financial Security Institute · Korean National Police Agency Korea University

cs.CR

Submitted: 2026-05-31

Updated: 2026-11-12

Comments: Accepted to ACSAC 2026

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Cyber threat intelligence (CTI) reports now serve as essential resources for capturing adversary tactics, techniques…

Key concepts

Schema-agnostic CTI Knowledge Graph Construction System
ANCHOR is a tool designed to create structured threat intelligence graphs without being tied to specific, rigid formats like STIX. It dynamically discovers and maps data elements onto formal ontology classes and properties, making it flexible for various threat intelligence standards.
Hybrid Ontology Discovery
This mechanism combines two methods: embedding-based semantic search to find relevant concepts and LLM-guided recursive navigation to precisely align those concepts with formal ontology structures. This hybrid approach ensures both broad discovery and high accuracy in type assignment.
SHACL-based Closed-Loop Correction
This is a validation process where every proposed data assignment is checked against the target schema using SHACL rules. If a violation occurs, the system attempts self-correction up to three times before re-invoking the discovery tool, ensuring structural integrity in the final knowledge graph.
Local LLM Performance Matching
ANCHOR achieves typing performance nearly identical to high-end enterprise LLMs when paired with smaller, locally deployed open-source models. This is possible because ANCHOR offloads complex structural reasoning to the formal schema rather than relying solely on the raw capabilities of a large external model.

Terminology

Summary

Cyber threat intelligence (CTI) reports now serve as essential resources for capturing adversary tactics, techniques, and procedures observed in modern attack campaigns While traditional CTI platforms reduce this intelligence to isolated indicators through fixed schemas such as STIX, ontology-based representations preserve the semantic relationships needed for structured threat analysis In this paper, we present ANCHOR, a schema-agnostic CTI knowledge graph construction system that bridges LLMs and formal ontology schemas Experimental results on the UCO, STIX, and MALOnt schemas show that ANCHOR outperforms existing baselines in ontology typing and schema compliance In addition, ANCHOR with a local LLM closely matches enterprise LLM typing performance, enabling privacy-preserving CTI analysis with high fidelity

Pruning redundant reasoning tokens

The paper introduces ANCHOR as a schema-agnostic CTI knowledge graph construction system that addresses three key challenges in existing approaches: (i) schemaspecific pipelines that require manual reconfiguration whenever the schema changes, (ii) prompt-based schema inclusion that fails to scale on large ontologies such as UCO, and (iii) reliance on enterprise LLM APIs that conflicts with privacy constraints when integrating sensitive internal incident data. The core of ANCHOR is hybrid ontology discovery, a search-and-navigate mechanism that dynamically explores largescale ontology schemas, combined with SHACL-based validation to enforce schema-compliant type assignments. p. 0

How the training signal is built

ANCHOR utilizes a hybrid ontology discovery mechanism that combines embedding-based semantic search with LLM-guided recursive navigation to align extracted elements to formal ontology classes or properties. The entity ontology typing phase employs Algorithm 1, which first performs an embedding-based search and then, if the confidence falls below a predefined threshold τentity, it activates hierarchical recursive navigation over root classes Uroots. p. 4 Similarly, predicate ontology typing extends this strategy using all three steps of Algorithm 1 to determine either an ObjectProperty URI or a DatatypeProperty URI. p. 5

Where the gains come from

The primary gains are demonstrated across three experimental results on the UCO, STIX, and MALOnt schemas. Specifically, ANCHOR achieves an average F1 of 0. p. 77371 for entity ontology typing and 0. p. 75484 for predicate ontology typing, outperforming baselines on all schemas. p. 7 Furthermore, the system demonstrates a schema non-compliance rate of 5. p. 82% on UCO, which is an improvement over CTINexus's 43.9%. When paired with locally deployed open-source LLMs like Qwen3. p. 15-35B, ANCHOR maintains 99. p. 62% (entity) and 97. p. 18% (predicate) of the best enterprise LLM's performance.

Schema Compliance and Privacy

The system ensures structural integrity through a SHACL-based closed-loop correction process during knowledge graph construction, where every candidate assignment is checked against the target schema before commitment. This validation mechanism allows ANCHOR to re-invoke the discovery tool when it detects a violation, limiting this self-correction to three retries. p. 9 The local deployment configuration directly addresses privacy regulations by allowing sensitive data integration without exposing it to enterprise LLM APIs. p. 9

Ablation Study Insights

The ablation study reveals that embedding-based search and hierarchical recursive navigation are fundamentally complementary, where embedding search drives entity ontology typing and recursive navigation drives predicate ontology typing. The analysis shows that the hybrid configuration improves performance by 22. p. 82% over Recurse-Only on predicate ontology typing because structural reasoning is needed for predicates, which is handled by recursive navigation rather than lexical similarity. p. 8 Additionally, the performance of ANCHOR offloads structural reasoning from the LLM to the schema, which reduces the performance gap between local and enterprise backbones. p. 9

Conclusion

ANCHOR successfully addresses schema dependency, scalability issues on large ontologies, and privacy risks associated with enterprise LLMs. By introducing hybrid ontology discovery combined with SHACL-based validation, ANCHOR constructs an ontology-aligned threat knowledge graph that preserves semantic connectivity while maintaining high typing fidelity. p. 0 The results confirm that schema-grounded reasoning is more critical than raw LLM capability for this task. p. 9 The system's ability to retain near-enterprise performance with local LLMs validates its utility for privacy-preserving CTI analysis.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Schema-agnostic extraction pipeline: ANCHOR decouples extraction logic from specific ontology schemas, supporting arbitrary OWL/SHACL ontologies at runtime without manual reconfiguration. This allows different OWL/SHACL ontologies to be loaded without manually reconfiguring the pipeline.

  2. Hybrid ontology discovery mechanism: The system combines embedding-based semantic search with LLM-guided recursive navigation to retrieve only task-relevant ontologies, integrated with closed-loop SHACL validation. This enables schema-aware reasoning within large and hierarchical ontologies while keeping each query within the context window.

  3. Knowledge graph construction with SHACL validation: ANCHOR applies a SHACLbased validation mechanism that evaluates every entity against the node shapes defined in the target ontology to ensure structural integrity, which is crucial because LLM-based extraction often omits mandatory attributes.

  4. Local LLM feasibility for privacy preservation: ANCHOR maintains high typing fidelity when paired with locally deployed open-source backbones, retaining 99.2% (entity) and 97.8% (predicate) of the best enterprise LLM’s performance. This supports privacy-preserving CTI analysis without a loss of typing fidelity.

  5. Confidence-based typing: ANCHOR rejects mappings below specific confidence thresholds to prevent corruption, using τentity = 0.45 and τpredicate = 0.30 as the default configurations for all subsequent experiments. This prevents low-confidence guesses from corrupting the knowledge graph, which is essential to avoid misleading threat attribution in security operations.

  6. Dynamic retrieval of ontology fragments: The system uses dynamic exploration that retrieves only task-relevant ontology fragments on demand instead of including the entire schema in the LLM prompt, which solves the scalability problem for large ontologies like UCO.

Sources

Related papers