TL;DR: AI security tools that rely on alerts, logs, and natural-language summaries still struggle to reduce risk because they lack structured context, according to Uptycs. Ontology-driven reasoning changes the control point from downstream triage to upstream risk reduction, where exploitability, privilege chains, and cross-domain relationships can be queried before incidents form.
At a glance
What this is: NHI Mgmt Group's analysis, drawing on Uptycs's blog post: AI security becomes materially more useful when it reasons over a structured cybersecurity ontology rather than raw alerts and logs.
Why it matters: This matters because IAM, cloud, endpoint, and NHI teams need AI that can identify exploit paths and privilege relationships before they become incidents, not just summarize them after detection.
Context
The problem is not whether AI can summarise security data, but whether it can understand how identities, permissions, workloads, and systems relate across cloud, endpoint, container, and Kubernetes environments. Without that structure, AI stays reactive, dependent on alerts and limited to downstream workflows rather than risk reduction.
A cybersecurity ontology is a machine-readable model of entities and relationships that preserves context across domains. In practice, that makes AI reasoning more useful for cloud security guidance because exploitability often emerges from how assets and identities connect, not from any single event in isolation.
Key questions
Q: How should security teams use AI to triage insider-risk findings without overwhelming analysts?
A: Security teams should use AI as a prioritization layer, not a replacement for policy. The practical goal is to reduce noise, surface the highest-severity activity first, and give analysts concise context for faster decisions. That works best when the model is tuned to the organisation’s data flows and policies, so routine events are filtered while genuinely risky behaviour remains visible.
Q: Why do AI security tools fail when they depend only on logs and alerts?
A: Because logs and alerts rarely encode the full relationship between identities, permissions, workloads, and boundaries. AI can infer patterns from that data, but it cannot reliably reason about causality if the underlying structure is missing. The result is reactive analysis rather than preventive security.
Q: What should teams evaluate before adopting ontology-driven AI security?
A: Check whether the platform can model identities, permissions, workloads, and cross-domain relationships as queryable objects. Also test whether its conclusions can be validated against ground truth. If it cannot explain why a path is risky, the system is still behaving like a summariser, not a reasoning layer.
Q: What is the difference between AI-assisted security and AI-native security platforms?
A: AI-assisted security adds summarisation, prioritisation, or workflow help on top of existing controls, while AI-native security embeds reasoning into the control plane itself. The difference is whether AI helps people work faster or helps the platform identify and reduce risk before incidents occur.
Technical breakdown
Why unstructured security telemetry limits AI reasoning
Most AI security tools still consume alerts, logs, events, and incident narratives as if the raw text were the problem to solve. That creates probabilistic inference without reliable context, so the model can summarise activity but cannot always explain causality or attack path. The practical limitation is structural: if relationships are not explicit, the AI has to guess how one identity, workload, or permission relates to another. That is why many tools are good at triage but weak at identifying pre-attack conditions that never generate a clean alert. See the Cloud Security Guide.
Practical implication: Use AI only where the underlying telemetry is normalised into entities and relationships, not where it is asked to infer structure from noise.
How ontology plus SQL changes the control plane
In this model, the ontology is queryable with SQL, and AI reasons over query results rather than raw data. That matters because SQL provides deterministic access to known relationships, while the model interprets bounded outputs instead of open-ended telemetry. The architecture reduces hallucination risk, validates conclusions against ground truth, and scales across large environments without forcing every answer through a token-heavy prompt. This is less about interface choice and more about making AI behave like a controlled analytical layer over security data.
Practical implication: Ground AI-driven investigations in queryable environment state so conclusions can be checked before they drive response or remediation.
Why identity relationships are the deciding factor in upstream risk
The strongest use cases are not generic summarisation but detection of privilege chains, identity misuse, and cross-domain attack paths. When an ontology can show which identities assume roles, which permissions enable movement, and which workloads talk across boundaries, AI can identify risk before exploitation. That makes identity the connective tissue of the security model, especially in hybrid estates where cloud, endpoint, and container activity intersect with service accounts and delegated access. The value is not that AI sees more events, but that it sees the structure behind them.
Practical implication: Map identity and permission relationships explicitly if you want AI to prioritise exploitable paths instead of merely ranking alerts.
Threat narrative
Attacker objective: The attacker aims to exploit structural blind spots in identity and workload relationships so risky paths remain invisible until damage is already underway.
- Entry begins when attackers exploit machine-speed opportunities hidden inside cross-domain relationships rather than triggering a single obvious alert.
- Credential or privilege access follows when identity misuse, overbroad permissions, or latent trust paths can be traversed without structured detection.
- Impact occurs when the security team sees only downstream symptoms, while the attacker has already used the environment's own relationships to move and act.
Breaches seen in the wild
- DeepSeek database exposure 2025: An unauthenticated DeepSeek ClickHouse database exposed over a million log lines with plaintext chat history and API keys in 2025.
- 12,000 secrets in LLM training data: Truffle Security found 11,908 live API keys and passwords hard-coded in web pages captured by Common Crawl, a dataset used to train LLMs.
Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.
NHI Mgmt Group analysis
Ontology-driven AI security is a control-plane shift, not a triage upgrade. If AI only consumes alerts, it inherits the limitations of the systems that generated them. The more important question is whether AI can reason over entities, permissions, and relationships before incidents form. That shifts value from summarisation to prevention, which is the kind of capability security architecture actually needs.
Identity is the anchor point for any useful security ontology. Cross-domain reasoning only becomes operational when identities, roles, workloads, and permissions are modelled as connected objects rather than isolated records. That is why the identity bridge is genuine here, and why AI security maturity guidance matters when teams want AI to evaluate risk paths instead of just report activity.
Structured reasoning will separate durable security platforms from temporary AI augmentation. Tools that sit on top of alerts will continue to help with workflow, but they will not redefine the risk model. Platforms that embed AI into a queryable control plane can expose exploitability earlier and with less ambiguity. Practitioners should treat ontology coverage as a strategic selection criterion, not a technical nice-to-have.
SQL is not an implementation detail in AI-native security, it is a trust boundary. Queryable state gives teams a way to test whether an AI conclusion matches the environment rather than the model's guess. That matters when decisions involve remediation priority, identity exposure, or attack path collapse. The practical conclusion is simple: if the reasoning layer cannot be validated, it should not be steering security action.
Read next: Cloud Security Guide
What this signals
Ontology coverage will become a procurement signal for AI security. Teams will increasingly need to ask whether a platform can reason across identity, cloud, endpoint, and container relationships instead of only consuming alerts. That makes the presence of a structured environment model a buying criterion, not a design preference.
AI security programmes will need to distinguish workflow automation from structural risk reduction. If a platform only speeds up incident handling, it does not alter exposure. The next maturity step is to demand AI that can surface exploitable identity and privilege conditions before they are used.
Agentic and AI-native security tools will be judged on validated reasoning, not model fluency. The more security operations depend on machine reasoning, the more important it becomes to verify conclusions against known environment state. That is where queryability, determinism, and identity relationships become operational, not theoretical.
For practitioners
- Define the security ontology before expanding AI use Map identities, permissions, workloads, containers, and cloud assets as linked entities so AI can reason over relationships rather than unstructured text.
- Move AI analysis onto queryable ground truth Use SQL or equivalent deterministic queries to ground model output in known environment state before it informs triage or remediation.
- Prioritise identity and privilege paths as first-class risk objects Track which identities can assume which roles and which permission chains create lateral movement opportunities across domains.
- Test whether AI can explain exploitability, not just anomalies Ask for the path, the enabling relationship, and the reason a condition is risky before accepting a model output as actionable.
Key takeaways
- AI security is most effective when it reasons over structured relationships, because raw alerts alone do not expose exploit paths or identity dependencies.
- The article frames ontology as a way to move from downstream summarisation to upstream risk reduction, especially where privilege chains and cross-domain trust matter.
- For practitioners, the question is whether AI can validate why something is risky against ground truth before it influences triage or remediation.
Related NHI Mgmt Group resources: Cloud Security Guide · AI Security Maturity Model · SOAR and Security Automation Guide · Top 10 SOC Challenges
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Non-Human Identity Top 10 and MITRE ATT&CK address the attack and risk surface, while NIST CSF 2.0 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Non-Human Identity Top 10 | NHI-05 — Overprivileged NHI | The article centers identity and permission chains as the exploitable structure AI should expose. |
| NHI-04 — Insecure Authentication | AI reasoning must account for how identities authenticate and connect across domains. | |
| Recommendation — Model and review overprivileged service and workload identities before using AI to prioritise risk paths. Trace authentication relationships across cloud and workload identities so AI can flag risky trust paths. | ||
| NIST CSF 2.0 | ID.AM-01 — Physical devices and systems within the organisation are inventoried | A queryable ontology depends on knowing which assets and identities exist in scope. |
| PR.AA-05 — Access Permissions, Entitlements and Authorizations | The article repeatedly ties risk to permission chains and exploitable access relationships. | |
| Recommendation — Inventory the assets and identities that feed AI reasoning so environment state can be trusted. Audit permissions and authorisations as linked risk objects, not isolated entitlements. | ||
| MITRE ATT&CK | TA0006;TA0008 — Credential Access; Lateral Movement | The article's attack-path logic centers on identity misuse and movement across connected systems. |
| Recommendation — Map identity-driven attack paths to credential access and lateral movement techniques in your detections. | ||
Key terms
- Cybersecurity Ontology: A cybersecurity ontology is a machine-readable model that describes assets, identities, permissions, workloads, and their relationships. It gives AI a structured view of the environment so conclusions can be grounded in explicit context rather than inferred from disconnected logs or alerts.
- Ontology-Driven Reasoning: Ontology-driven reasoning is the use of structured relationships and queryable entity data to help AI infer risk, causality, and attack paths. It is more reliable than free-form prompt interpretation because the model reasons over defined constraints and known environment state.
- Upstream Risk Reduction: Upstream risk reduction means identifying and reducing exploitable conditions before they become active incidents. In security operations, this shifts attention from response speed alone to the earlier structural issues that make compromise possible in the first place.
- Security Control Plane: The security control plane is the management layer that defines, distributes, and enforces security policy across systems. It coordinates identity, access, telemetry, and response decisions, often across cloud, endpoint, network, and application domains. In practice, it is where authorization rules, trust signals, and enforcement actions are centrally governed.
Deepen your knowledge
NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, machine identity security, and secrets management. It helps security practitioners connect identity control to the broader security architecture their programmes depend on.
Published by the NHIMG editorial team on October 22, 2026.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org