RESPONSIBLE AI

AI that strengthens security decisions without bypassing human control .

Hayrok uses AI to help security teams understand evidence, investigate findings, analyze attack paths, prioritize remediation, and summarize risk. AI in cybersecurity must respect authorization, understand evidence boundaries, communicate uncertainty, and remain governed — never turning a recommendation into a high-impact action without the appropriate controls.

RESPONSIBLE AI PRINCIPLESCORE RESPONSIBLE AI STATEMENT
01 · OUR COMMITMENT

Intelligence should increase confidence, not create new uncertainty

Speed alone is not enough. Security decisions affect production systems, customer data, identity, cloud, controls, and validation itself. Hayrok therefore treats AI as a governed decision-support capability rather than an unrestricted authority — organized around seven principles.

◆01
Evidence before assertion
Claims are proportional to what the validated record can support.
✋02
Human accountability
AI supports decisions; people remain responsible for consequential ones.
§03
Authorization before access
AI cannot become a shortcut around identity, tenant, or role permissions.
⚙04
Governance before action
Recommendations do not automatically become executed actions.
?05
Transparency about uncertainty
Confirmed, observed, inferred, unknown — each labeled clearly.
⌘06
Privacy & security by design
AI infrastructure treated as part of the security boundary itself.
↻07
Continuous evaluation
Ongoing red-team testing, benchmarks, and monitoring — not just launch checks.
02 · EVIDENCE BEFORE ASSERTION

AI answers should be grounded in what Hayrok can actually prove

Nectar reasons over the validated record. Every response distinguishes what the evidence proves from what it merely suggests — and is comfortable saying it doesn't know. Click each verdict to see how it reads.

CONFIRMED
The record proves it
#166534
OBSERVED
Relevant, not decisive
#92400E
INFERRED
Contextual reasoning
#9A3412
UNKNOWN
Not enough evidence
#475569
CONFIRMED Was the finding exploited in production?
"Exploitation is confirmed. Request/response evidence shows an authenticated user retrieved records outside their tenant boundary; runtime and control-response artifacts support the conclusion."
Even the correct refusal to answer is a security signal. Nectar is designed to say when it can't.
03 · GOVERNANCE BEFORE ACTION

Nectar reasons. Genesis executes. Humans decide between them.

Hayrok separates intelligence from adversarial execution. A recommendation is never the launch.

NECTAR
Evidence-grounded intelligence
Reasons. Explains. Investigates. Recommends. Summarizes. Guides revalidation. Never launches.
HUMAN
Decision & authorization
A human reviews the recommendation and — where required — approves the action.
GENESIS
Governed orchestration
Plans. Applies policy. Enforces scope. Requests approval. Executes safely. Preserves evidence.
Nectar may recommend "revalidate this finding" — the recommendation itself does not launch the validation. Authorization, entitlement, safety, policy, and approval controls still apply.
04 · WHAT NECTAR CAN & CANNOT DO

Two clear lists. No ambiguity.

✓Explain findings
✓Interpret evidence
✓Analyze attack paths
✓Summarize defensive response
✓Identify detection gaps
✓Recommend remediation
✓Prepare executive summaries
✓Guide revalidation
✓Compare validation outcomes
✓Identify missing evidence
05
SECTION 05

Privacy by design

Customer data is used only for authorized purposes. Access is limited to what an authorized workflow requires.

Tenant isolation
Purpose limitation
Least-privilege access
Data minimization
Access logging
Retention controls
Sensitive-data protection
Authorized retrieval
06
SECTION 06

Security by design

AI infrastructure is treated as production security infrastructure — not an isolated experimental feature.

Auth & authorization on every request
Input validation on untrusted content
Output evaluated before governed actions
Secrets protection from model context
Controlled network pathways
Logging & monitoring
Provider security review
Change-controlled model rollout
07
SECTION 07

Prompt injection and untrusted content

Security platforms process untrusted information. Retrieved content is treated as evidence, not authoritative instructions.

Prompt injection controls
Indirect prompt injection controls
Instruction hijacking defenses
Tool manipulation limits
Data exfiltration guardrails
System policy remains authoritative
08
SECTION 08

Tool use must be governed

Where AI can interact with tools, access is explicitly scoped. AI cannot grant itself permissions or expand its own reach.

User authorization
Tenant & environment scope
Requested action allowlist
Resource risk level
Validation objective binding
Policy & approval status
09
SECTION 09

High-impact actions require additional control

The higher the impact, the stronger the governance. AI recommendation alone is never sufficient authorization.

Explicit user confirmation
Policy reevaluation
Human approval
Step-up authorization
Time-limited authorization
Detailed audit records
10
SECTION 10

Safe failure

AI does not fill missing security evidence with speculation. Prefer insufficient_evidence over false confidence.

insufficient_evidence
unknown
inconclusive
telemetry_unavailable
validation_required
Detection status cannot be determined
11
SECTION 11

AI should not determine security truth alone

Deterministic systems remain authoritative where they can be. AI assists with interpretation, summarization, and reasoning.

Authentication
Authorization & entitlements
Policy enforcement
Quotas
Hard validation rules
Required evidence checks
Cryptographic verification
12
SECTION 12

AI and evidence integrity

AI-generated summaries are derived content. They never replace or rewrite the underlying evidence record.

No modifying original evidence
No rewriting timestamps
No changing validation results
No altering provenance
No deleting contradictory evidence
No silent reclassification
13
SECTION 13

AI and autonomous validation agents

Autonomy stays bounded. Every agent runs with defined purpose, tools, tenant context, scope, and stop conditions.

Recon Agent
Planner Agent
Exploit Agent
Validation Agent
Evidence Agent
Reporting Agent
No agent has full-lifecycle authority
14
SECTION 14

Separation of duties

Hayrok avoids architectures where one autonomous component can select, approve, execute, judge, suppress, and close.

Planning
Policy evaluation
Human approval where required
Execution
Independent validation
Evidence preservation
Finding recorded
15
SECTION 15

Model & provider governance

Models are governed as critical service dependencies. New providers undergo security and privacy review before production.

Security & privacy review
Data handling & retention
Data residency
Model quality benchmarks
Tool-use behavior
Prompt-injection resilience
Contractual protections
16
SECTION 16

Continuous evaluation & red-teaming

AI is evaluated continuously — grounding, citations, authorization, tenant isolation, prompt injection, tool safety, uncertainty, and recommendation quality.

Grounding
Citation accuracy
Hallucination probes
Authorization boundary tests
Tenant-isolation checks
Prompt-injection red-team
Tool-safety evaluation
Uncertainty acknowledgment
17
SECTION 17

Monitoring & abuse detection

Security-relevant AI behavior is observable — unauthorized-access attempts, prompt manipulation, unusual tool requests, policy failures.

Repeated unauthorized-access attempts
Prompt manipulation patterns
Sensitive-data requests
Suspicious tool requests
Excessive query volume
Policy failures
18
SECTION 18

Auditability

Relevant Nectar interactions record metadata proportionate to the security sensitivity of the workflow.

User
Tenant
Timestamp
Requested operation
Context sources
Evidence references
Recommendation
Resulting user action
19
SECTION 19

Customer control

Enterprise customers get visibility and choice. Only currently available capabilities are presented as generally available.

Enable/disable AI features
Role-based AI access
Policy-based AI access
Configurable retention
Model-provider restrictions
AI audit logs
Approved-use policies
EX
SECTION 29

Responsible AI in practice

◆NectarFINDING · HAY-2026-1041
You
Why should we prioritize this finding?
◆
CONFIRMED
Exploitation is confirmed in production, the affected API is internet-accessible, the application authorization control failed, no correlated detection was observed, and the validated path reaches the Customer Database.
EVIDENCE USED
[1] API exposure [2] AuthZ validation [3] Runtime presence [4] App control response [5] Detection result [6] Validated path [7] Crown-jewel class
Uncertainty: current evidence confirms unauthorized record access but does not establish bulk data extraction.
Recommended action: fix tenant-level authorization, evaluate equivalent endpoints, add detection for cross-tenant access, and revalidate after deployment.
USER CHOICEPrepare Revalidation → Genesis handles authorization, policy, approval, and controlled execution.
30
SECTION 30

Twelve commitments — in one page

01Evidence before assertion 02Human accountability 03Authorization before access 04Governance before action 05Transparency about uncertainty 06Privacy by design 07Security by design 08Least privilege 09Separation of duties 10Traceability 11Safe failure 12Continuous evaluation
34
SECTION 34

Responsible AI FAQ

Does Nectar make security decisions for customers?+
Nectar is designed as a decision-support and intelligence capability. Consequential actions remain subject to user authorization, platform policy, and approval workflows.
Can Nectar launch a validation?+
Nectar may recommend or prepare a validation, but Genesis owns governed orchestration and execution. Applicable authorization and approval controls remain in effect.
Does Nectar have access to everything in a tenant?+
No. Access is constrained by the requesting user's permissions and the relevant tenant, organization, resource, and evidence-access policies.
Can Nectar access another customer's data?+
The Hayrok architecture is designed to enforce tenant isolation across AI retrieval and platform access.
Does Nectar always know whether a finding is exploitable?+
No. Nectar relies on available validation evidence. If exploitability has not been established, it clearly communicates that status.
How does Hayrok reduce hallucinations?+
Evidence-grounded retrieval, explicit evidence states, source references, deterministic controls where appropriate, and continuous evaluation. No generative AI is represented as incapable of error.
Can Nectar override a policy decision?+
No. Policy systems remain authoritative.
Is customer data used to train AI models?+
Hayrok does not use Customer Data to train general-purpose AI models. Model-provider configuration and contractual practices are reflected in the Trust Center and subprocessor documentation.
Can customers disable AI?+
Enterprise plans include options to enable or disable specific AI features and constrain AI access via role- and policy-based controls. Contact Hayrok for the options available on your plan.
Does Hayrok use multiple AI model providers?+
Where multiple providers are used, current providers and processing arrangements are reflected in Hayrok subprocessor and trust documentation.

Evidence before assertion. Governance before action. Humans remain accountable.

Hayrok uses AI to help security teams understand complex validation evidence and make better decisions — without weakening the controls that protect their environments.