Cybersecurity was not literally designed before artificial intelligence existed, but many security environments were built around fragmented logs, isolated tools and limited data sharing. That is the issue raised in an Inside Cybersecurity conversation between Cylake founder and CEO Nir Zuk and René Bonvanie, summarized by Cybercrime Magazine on August 31, 2026. The argument is straightforward: an AI feature cannot produce reliable security decisions from data it cannot access, cannot interpret or cannot trust.
For security leaders, the practical lesson is larger than buying an AI function. An organization needs a governed, sufficiently complete view of identities, assets, activity, vulnerabilities, controls and business context. Without that foundation, AI may accelerate analysis of a narrow slice of reality while leaving the most important blind spots untouched.
As an Amazon Associate I earn from qualifying purchases.
What the Cylake argument actually says
The Cylake discussion frames AI usefulness as a data-access problem. Nir Zuk’s position, as summarized in the August 31, 2026 article, is that the quality of an AI system’s security output depends on the data available to it. If telemetry is missing, stale, trapped in separate systems or stripped of context, an otherwise capable model has less to work with.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
This is an architectural argument, not proof that every established security platform has failed. The summary does not present a benchmark, breach-rate comparison or independent test showing that existing products are universally inadequate. It identifies a recurring constraint: security teams often operate with fragmented data architectures, while AI needs connected and appropriately governed information to reason across events.
#1 Best Overall
An AI feature is not the same as an AI-ready security program
A vendor can add natural-language search, alert summarization or automated response to a product without changing the underlying data estate. That feature may still be useful inside its permitted boundary. It does not automatically gain visibility into cloud identities, endpoint state, software inventories, third-party access, physical systems or business priorities held elsewhere.
AI readiness therefore has two layers. The first is the model or feature: what it can classify, generate, correlate or automate. The second is the information architecture: which records are available, how they are normalized, how quickly they arrive, who can use them and how their quality is checked. The second layer determines how far the first can safely reach.
Why fragmented security data limits AI
Silos hide relationships
Security events rarely explain themselves. A suspicious login may only become meaningful when matched with an unfamiliar device, a new privilege assignment, a vulnerable application and an unusual data transfer. Those facts commonly live in different systems owned by different teams. If an AI workflow sees only the login record, it may rank the event incorrectly or miss the relationship entirely.
Inconsistent formats create false uncertainty
Different tools identify the same user, device or application in different ways. One system may use an email address, another an employee number and a third a cloud-specific identifier. Hostnames can change, timestamps can use different time zones and severity labels can mean different things. Before an AI system can correlate records, the organization has to establish whether those records refer to the same entity and whether their time and confidence fields are comparable.
Missing history weakens detection
Many investigations depend on a baseline: what is normal for this account, workload or network segment? A model that receives only a short retention window cannot reliably distinguish a rare legitimate action from a new attack technique. Deleting or isolating historical data may reduce storage costs, but it also removes evidence needed for trend analysis and incident reconstruction.
Context determines priority
An alert involving a development server does not necessarily deserve the same response as an equivalent alert involving a payment system or an administrator account. Business criticality, regulatory obligations, maintenance schedules and approved exceptions supply that context. If those attributes are absent, an AI system may optimize for technical severity rather than actual organizational risk.
Access restrictions can create blind spots
Good security architecture cannot simply expose every record to every model. Privacy, contractual limits, segmentation and least-privilege rules still apply. The challenge is to make the right data available to the right analytic function under a documented purpose, rather than treating unrestricted access as the definition of visibility.
Generative AI and machine learning are not interchangeable
The Cylake summary indicates that the conversation distinguishes generative AI from machine learning, but it does not provide a transcript detailed enough to report every distinction made by the speakers. In practical terms, readers should avoid using the labels as synonyms.
Machine-learning systems are commonly used to classify, score or detect patterns in structured or behavioral data. Generative AI produces new text, code or other content and can provide an interface for investigation or assistance. A generative assistant may explain an alert, while a detection model may assign a risk score; either can be wrong when the underlying records are incomplete. The governance, testing and access controls should be appropriate to the capability being used, not to a marketing label.
What NIST guidance adds to the discussion
AI RMF and the Generative AI Profile
NIST describes the Artificial Intelligence Risk Management Framework (AI RMF) 1.0 as a voluntary resource for organizations designing, developing, deploying or using AI. Its Generative AI Profile is a companion resource focused on risks that are novel to, or worsened by, generative AI. The profile organizes suggested actions around governing, mapping, measuring and managing those risks.
Rank #3
NIST published the Generative AI Profile on July 26, 2024, and the page was updated on April 8, 2026. It is cross-sector guidance rather than a mandatory regulation. Organizations can use it to structure questions about data provenance, evaluation, security, privacy, human oversight and accountability, while adapting the controls to their own legal and operational environment.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallThe Cyber AI Profile is still being developed
NIST’s Cybersecurity Framework Profile for Artificial Intelligence should not be described as a finalized standard based on the 2026 material available here. NIST IR 8596 was an initial preliminary draft, with public comments closing on January 30, 2026. Its stated purpose covers both cybersecurity risks related to AI systems and opportunities to use AI to improve cyber capabilities.
Workshop reports identified continuing work on governance challenges, attack surfaces, taxonomy, risk-based guidance, usability and AI-enabled defense. NIST reports from August 2026 show that profile development and discussion were continuing. This status matters: teams can use the draft and workshop themes as inputs, but should not claim compliance with a completed NIST Cyber AI standard.
A practical blueprint for making security data usable by AI
The following sequence translates the data-visibility concern into an implementation program. It is practical guidance, not a claim that NIST mandates these exact steps.
- Define the decisions AI is expected to support. Specify whether the system will triage alerts, investigate identities, prioritize vulnerabilities, recommend containment or answer analyst questions. Each use case has different data, latency, accuracy and approval requirements.
- Inventory the sources behind each decision. List identity providers, endpoint and network telemetry, cloud audit records, vulnerability information, application inventories, ticketing systems and business-asset registers. Record the owner, retention period, update frequency and known gaps for every source.
- Establish common entities and identifiers. Create authoritative mappings for people, service accounts, devices, workloads, applications and locations. Preserve source identifiers so an analyst can trace an AI conclusion back to the original record.
- Normalize time, severity and confidence. Document time zones, clock synchronization, event schemas and severity definitions. Keep the original value alongside any transformed value; normalization should improve correlation without destroying evidence.
- Attach business and operational context. Add asset criticality, data classification, ownership, maintenance windows, approved exceptions and recovery priorities. Require an owner to review context that can materially change a response recommendation.
- Apply purpose-based access controls. Decide which fields each analytic function may read, which data must be masked and which actions require a human approval. Log model access and preserve an audit trail for generated recommendations and automated changes.
- Measure data quality before model quality. Track coverage, freshness, duplication, conflicting identities, missing fields and delivery failures. A high model score on a narrow, clean sample does not demonstrate that production visibility is sufficient.
- Test with realistic failure cases. Deliberately remove a telemetry source, introduce stale asset ownership or replay an identity change. Check whether the system signals uncertainty, degrades safely and gives analysts a way to verify the result.
- Monitor drift and revise the data contract. New cloud services, reorganized teams and changed logging settings can invalidate old mappings. Set owners and review intervals for schemas, retention, permissions and model behavior.
Use this data-readiness checklist
| Layer | Questions to answer | Warning sign |
|---|---|---|
| Coverage | Are the identities, assets, workloads and telemetry relevant to the use case represented? | The model is confident while major environments are absent. |
| Freshness | How quickly do records arrive, and when was each asset or permission last verified? | Recommendations rely on stale ownership or outdated exposure data. |
| Consistency | Can records from separate tools be joined using stable identifiers and shared time semantics? | The same user or device appears as multiple unrelated entities. |
| Context | Does the system know business criticality, approved changes and data sensitivity? | Technical severity is treated as the only measure of risk. |
| Provenance | Can an analyst inspect the original event and transformation steps? | A generated explanation cannot be traced to evidence. |
| Governance | Are purpose, access, retention, review and human-approval rules documented? | Teams cannot explain who may use the data or why an action occurred. |
How to evaluate an AI security capability
Procurement should begin with the information boundary, not the demonstration. Ask the provider and internal owners to document the following criteria:
- Security-data coverage: Which identity, endpoint, cloud, network, application and vulnerability sources are supported, and what integration effort is required?
- Asset and identity resolution: How are duplicate accounts, ephemeral workloads, renamed devices and service identities handled?
- Data protection: Where are prompts, telemetry and derived records stored? What masking, encryption, retention and tenant-isolation controls apply?
- Access and action limits: Can permissions be scoped by role, field, environment and purpose? Which responses are advisory, and which can change systems?
- Evidence and explainability: Does each recommendation show the events, timestamps and transformations supporting it?
- Evaluation: What is measured for false positives, false negatives, latency, drift and abstention when evidence is insufficient?
- Operational fit: Can analysts correct an entity mapping, provide feedback, override an action and export records for an investigation?
These are evaluation axes, not a comparison of named products. A tool with many connectors may still be unsuitable if the organization cannot govern the resulting data flow, while a narrower system may be appropriate for a tightly defined use case with strong evidence and human review.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Common failure modes and how to recover
Starting with a chatbot
A conversational interface is visible and easy to demonstrate, so teams may deploy it before defining decisions, evidence and approval boundaries. Recover by selecting one measurable workflow, such as analyst triage, and documenting the records required for a correct answer.
Connecting everything without ownership
More feeds do not automatically mean better visibility. Unowned, duplicated or contradictory data can increase noise and expose sensitive fields unnecessarily. Assign a data owner and quality thresholds to each source before expanding the ingestion scope.
Automating irreversible actions too early
Containment, account suspension and configuration changes can disrupt legitimate work when context is missing. Begin with recommendations or reversible actions, require explicit approval for high-impact changes and retain the evidence used to make the recommendation.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteTreating a draft as a certification checklist
NIST’s Cyber AI Profile work remains in development in the records described above. Use preliminary material to inform risk discussions, but do not claim that an organization or product is certified against a finalized profile that has not been issued.
Best Value
Confusing fluent output with verified insight
Generative systems can produce a plausible explanation even when the relevant event is absent. Require citations to internal evidence, confidence or uncertainty indicators and a direct path for an analyst to inspect source records.
What the AI-era warning means for security leadership
The central decision is architectural: will AI receive a coherent, governed view of the environment, or only disconnected fragments? Leaders should fund data ownership, identity resolution, logging quality and access governance alongside model experimentation. They should also set a boundary between assistance and authority, with humans accountable for consequential decisions.
NIST’s voluntary AI RMF and Generative AI Profile provide a way to organize governance, mapping, measurement and management questions. The evolving Cyber AI Profile work signals that both protecting AI systems and using AI for defense are active areas of standardization, not settled checkboxes.
Cylake’s warning is therefore best read as a design test. If an organization cannot explain what data an AI system can see, how current and reliable that data is, what it cannot see and who approves its actions, the limiting factor is not necessarily the model. It is the security architecture around the model.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
