Access control
Access is granted on least privilege; administrator accounts use two-factor authentication; access is revoked as soon as a person leaves or a supplier contract ends.
Security & AI governance · v1.0
Information Security and AI Governance Policy
We put AI to work on our customers’ systems and data, so we publish the rules first: every release is approved by a person, every AI run is logged, and credentials never reach an AI model. Our management system uses ISO/IEC 27001 and ISO/IEC 42001 as its framework. We currently plan to obtain certification within 12 to 18 months.
- Policy version
- v1.0 · effective 2026-10-18
- Reference frameworks
- ISO/IEC 27001:2022 · ISO/IEC 42001:2023
- Third-party certification
- Not yet obtained
- Internal audit capability
- Staff qualified as ISO/IEC 42001 AIMS internal auditors
- Security reports
- [email protected]
Security & AI governance
Least privilege
Only the access a job needs.
Human in the loop
A person releases every change.
Traceable
Every AI run and change is logged.
Independent review
No one reviews their own output.
§1
Scope and framework
This policy covers Heyu Applied Intelligence Co., Ltd. and every PVL.AI website, platform, and customer project — the code, cloud services, data, and the people and AI agents that work on them.
We build our management system on ISO/IEC 27001:2022 (information security) and ISO/IEC 42001:2023 (AI management). We do not yet hold third-party certification to either standard; we will run the system and build up its records. We plan to obtain formal certification within 12 to 18 months of this policy taking effect.
§2
Information security controls
Credentials and keys
Credentials and API keys live in a managed cloud secret store, never in code. Repositories run secret scanning with push protection, and we are establishing a key-rotation process.
Change management
Every production change goes through a pull request and automated checks. Deployments that publish externally need the owner’s explicit approval at the time, and are logged.
Customer data isolation
Each customer project has its own repository. Multi-tenant services isolate tenants with database-level access rules, verified by automated tests.
Backup and continuity
Recovery targets are set in three tiers by service importance (see the table below). A backup is treated as valid only after a successful restore drill.
Supplier management
Cloud and AI-model suppliers will be kept in a supplier register recording the data they handle, where it is stored, and their data-retention terms.
| Service tier | Max data loss (RPO) | Max recovery time (RTO) |
|---|---|---|
| Tier 1 · critical | 1 hour | 2 hours |
| Tier 2 · standard | 6 hours | 4 hours |
| Tier 3 · other | 24 hours | 24 hours |
§3
Data classification and the limits of AI use
Data falls into four classes — public, internal, customer confidential, and credentials. The principle for each is set out below; the detailed classification rules and supplier register are being built.
Before adopting an AI-model supplier, we confirm its data-retention terms and whether it trains on submitted data.
ClassExamplesMay AI process it?
- L1PublicWebsite content, published articlesYes
- L2InternalInternal documents, plans, codeOnly with model suppliers that have been assessed and registered
- L3Customer confidentialData and documents customers provideOnly as the contract allows, with the supplier recorded
- L4CredentialsPasswords, keys, tokensNever
§4
AI governance
01
AI agent produces
code or content
02
Independent review
different model family
03
Human approval
owner signs off
04
Human releases
AI cannot deploy
Human in the loop
Publishing and production deployment are always carried out by a person. AI agents may prepare a change but cannot release it.
Traceability
Every AI-agent run is logged, and code changes written with AI assistance name the model involved.
Independent review
AI-produced content and code are never reviewed by the agent that produced them; the reviewer uses a different model family, and a person approves.
Pre-release evaluation
Customer-facing AI services will be tested against fixed test sets for answer accuracy, citations, and safe behaviour before release. An inconclusive result blocks release.
Controlled execution
AI development tools run in an isolated sandbox with network access limited to an allow-list. The sandbox an AI agent can reach holds no production credentials; anything that must touch production is run by a person through a separately approved path. AI-agent actions are graded into four risk levels; any action that writes to production must be approved by a person each time, with the approver recorded, or the system refuses to run it.
Tamper-evident audit trail
Approval, block, and usage records will be chained by hash, so altering or deleting any entry breaks the chain at verification.
Usage and budget limits
Every AI service will have usage and spending limits, with every call metered; when a limit is reached the default is to stop, not to carry on.
Approved model list
We are building an approved list of model suppliers; once in place, only suppliers that have been assessed and listed will be used. Adding or changing a model first requires a data-terms check and pre-release evaluation.
§5
Incident reporting and contact
To report a security weakness or a suspected data leak in a PVL.AI service, email [email protected] with “Security” in the subject line. We confirm, contain, and record a post-incident review; incidents affecting customer data are notified to affected customers as law and contract require.
Email a security report§6
Review and updates
This policy is approved by the company’s responsible officer and reviewed at least once a year, or sooner when the organisation, services, or regulations change materially. Each update is marked on this page with its version and effective date.
Internal audits are carried out by staff qualified as ISO/IEC 42001:2023 AI management system (AIMS) internal auditors, and their findings feed the annual management review.
Is PVL.AI certified to ISO 27001 or ISO 42001?
Not yet. Heyu Applied Intelligence Co., Ltd. (PVL.AI) builds its information-security and AI management system on ISO/IEC 27001:2022 and ISO/IEC 42001:2023 — covering access control, credential management, change management, customer data isolation, data-classification principles, and AI rules for human oversight, traceability, and independent review — and plans to obtain formal certification within 12 to 18 months of this policy taking effect.
Need more security detail?
For a partnership review we can walk you through our data flows, the cloud and AI suppliers we use, and the corresponding controls.