They were built to store and query human-readable events, not to supply structured context for machine reasoning. When enrichment happens too late or data is fragmented across brittle integrations, AI produces weaker output and more operational cost. AI-native design moves context earlier and standardises access paths.
Why This Matters for Security Teams
Legacy SIEM and telemetry stacks were optimised for alerting, audit trails, and after-the-fact investigation. generative ai changes the requirement: the system now needs structured, low-friction context that can be consumed by models at runtime, not just searchable logs for analysts later. When enrichment is delayed, normalised poorly, or scattered across point integrations, AI outputs become inconsistent and expensive to produce. That gap shows up clearly in incidents like the DeepSeek breach, where exposed secrets and weak operational boundaries amplified downstream risk.
Current guidance suggests that security telemetry should support machine reasoning as well as human review, especially where AI systems are expected to summarise, correlate, or respond autonomously. NIST’s NIST AI 600-1 GenAI Profile reinforces the need for context quality, provenance, and governance around AI inputs and outputs. In practice, many security teams discover that their telemetry architecture is fragmented only after an AI use case fails to retrieve the right context fast enough to matter.
How It Works in Practice
Generative AI depends on context packaging, not just event retention. That means the telemetry layer must expose data in forms a model can consume safely and consistently: incident metadata, asset identity, identity-to-resource relationships, policy state, and recent control actions. If those signals remain buried in separate tools, the model has to infer across brittle joins, which increases latency and hallucination risk.
Practitioners usually need to shift from log-first thinking to context-first design:
- Normalise event schemas early so AI can reason across sources without custom parsing for every query.
- Attach identity, asset, and policy context at ingestion time rather than during investigation.
- Use well-defined retrieval paths so the model sees the same authoritative state that analysts do.
- Preserve provenance and timestamps so AI responses can be traced back to source evidence.
- Separate human alerting workflows from machine-consumption pipelines where necessary.
This is where the difference between generic observability and AI-ready telemetry becomes obvious. SIEM products often excel at detection and search, while AI systems need structured context, consistent semantics, and predictable access patterns. Controls from NIST SP 800-53 Rev 5 Security and Privacy Controls help define governance and logging expectations, but they do not by themselves solve the data-shaping problem for AI consumption. NHIMG’s analysis of Sumo Logic Breach shows why telemetry trust, access control, and data handling cannot be treated as separate concerns. These controls tend to break down when telemetry is collected through dozens of disconnected integrations because schema drift and inconsistent enrichment make the context unreliable.
Common Variations and Edge Cases
Tighter context pipelines often increase integration overhead, requiring organisations to balance AI readiness against operational complexity. That tradeoff is real: not every telemetry source needs to be rebuilt, and not every use case justifies full semantic normalisation.
Best practice is evolving for a few common edge cases. High-volume security operations may keep raw log storage in the SIEM while routing a curated subset of enriched signals to AI systems. Regulated environments may need stricter separation between investigative data and model prompts. Some teams also discover that AI works well on incident narratives but poorly on noisy, low-quality detection streams, so the right answer is selective context curation rather than universal ingestion.
There is no universal standard for this yet, but the direction is clear: AI should not be forced to infer meaning from fragmented telemetry when that meaning can be standardised upstream. The more the environment depends on brittle parsers, late enrichment, and tool-specific joins, the more likely the AI layer is to underperform or misstate the state of the system. A legacy stack becomes a bottleneck when it cannot supply trusted context at the speed of model inference.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10, CSA MAESTRO and OWASP Non-Human Identity Top 10 address the attack and risk surface, while NIST CSF 2.0 and NIST AI RMF set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | DE.CM-1 | Telemetry quality and continuous monitoring are central to AI-ready security data. |
| NIST AI RMF | GOVERN | AI governance requires trusted context, provenance, and accountability for inputs. |
| OWASP Agentic AI Top 10 | A3 | AI systems fail when context is weak or retrieval is unreliable. |
| CSA MAESTRO | DAG-02 | Agentic workflows depend on reliable data paths and runtime context. |
| OWASP Non-Human Identity Top 10 | NHI-06 | Telemetry stacks often expose or mishandle secrets and identity context. |
Reduce secret exposure in logs and ensure NHI signals are handled with least privilege.
