SDAIA confirmed 48 PDPL violations in 2025. Every National ID, Iqama or IBAN pasted into a public AI tool is processing you cannot see or prove. SentraGuard finds it, governs it and proves it, inside the Kingdom.
Find your shadow AI on the day you deploy. Reach full shadow AI governance in one week.
ISO 27001 CERTIFIED
Sensitive data does not leave through files any more. It leaves in a prompt, a paste and an agent's tool call, usually on a personal account. Usage control decides what happens at that moment.
67% of employees reach AI from non-corporate accounts on corporate devices. No log, no attribution, no answer for the auditor. Verizon DBIR 2026
Shadow AI is now the third most common non-malicious insider action in DLP data, with detections up 4x in a year. Source code is the top data type uploaded. Verizon DBIR 2026
Coding agents and MCP servers call tools and write to data. A tool list cannot govern a tool call.
Disclosing sensitive data can cost up to SAR 3 million and two years in prison. Other violations cost up to SAR 5 million, doubled for repeat offenders. SDAIA confirmed 48 violations in 2025 alone. DLA Piper · SPA 2026
The session blind spot
SDAIA, the NCA and SAMA each set rules that reach AI use at work. Here is what applies, and the evidence SentraGuard produces for each.
The PDPL has applied in full since 14 September 2024. Breaches go to SDAIA within 72 hours. Controllers register on the National Data Governance Platform, and some must appoint a DPO. Transfers to countries without adequacy need safeguards under the Transfer Regulations of September 2024.
ECC-2:2024 and the Cloud Cybersecurity Controls CCC-2:2024 carry data localisation provisions. The Data Cybersecurity Controls DCC-1:2022 cover the data lifecycle. The NCA AI Cybersecurity Guidelines, in consultation until 5 August 2026, cover generative and agentic AI and third-party AI.
The framework's third-party section covers outsourcing and cloud computing. Material outsourcing needs SAMA approval first. A public AI tool that holds customer data is a third party nobody approved.
SDAIA published Generative AI Guidelines for Government and for the Public in January 2024, the AI Ethics Principles in 2023 and the AI Adoption Framework in 2024. The Kingdom has no AI-specific law yet. A draft Global AI Hub Law went to consultation in 2025.
Checked 24 September 2026. Regulations change; confirm current text with your counsel.
The browser plugin sees web apps and extensions. The IDE extension sees Cursor, VS Code and Copilot. The SDK and API gateway see your own apps. The endpoint agent runs on Windows, macOS and Linux and sees desktop AI apps, CLI agents, local models and MCP servers. Everything lands in one console.
Every surface reports to the same API-first backend, so one policy applies everywhere and every event lands in the same log.
A bank can run the whole pilot on the first rung and see every session, account and data class with no user impact. Each rung above adds control for one app, one group or one data class at a time.
Mark each app sanctioned, tolerated or unsanctioned. New apps inherit their category's decision until you change it. Then set the action by user group and by the data class detected in the prompt.
| Application | Retail ops | Engineering | Contractors |
|---|---|---|---|
| ChatGPT Enterprisesanctioned · SSO | Allow | Allow | Redact |
| ChatGPT personalunsanctioned | Coach | Coach | Block |
| Claude Codetolerated · engineering | Block | Redact | Block |
| DeepSeek webunsanctioned · trains by default | Block | Block | Block |
| M365 Copilotsanctioned | Allow | Allow | Coach |
A banner in the session, not a ticket the next day. The user sees why, what to do instead, and what is logged. Most usage moves to the approved tool without a block.
Illustrative rules. Your policy sets the real ones, per app, per group and per data class.
Every decision is logged with user, app, data class and action, and streamed to your SIEM.
Unsanctioned app on a personal account. The user sees the approved assistant and can continue; the session is logged.
Detection runs on the session, on every surface, with the same data classes. Redact keeps the workflow moving. Block stops the ones that must not leave.
Inventory, attribution, policy log, exceptions, a regulator clause map and a signed export. Produced from the same events the SIEM receives, so the auditor and the SOC read one truth.
Select a jurisdiction. Each dot is an artifact SentraGuard produces for that framework.
| Framework | Inventory | Attribution | Policy log | Redaction proof | Clause map | Signed export |
|---|---|---|---|---|---|---|
| SAMA Cyber Security Framework | ||||||
| NCA Essential Cybersecurity Controls, ECC-2:2024 | ||||||
| NCA Cloud Cybersecurity Controls, CCC-2:2024 | ||||||
| NCA Data Cybersecurity Controls, DCC-1:2022 | ||||||
| NCA AI Cybersecurity Guidelines (draft, 2026) | ||||||
| PDPL, Royal Decree M/19, and its regulations | ||||||
| SDAIA Generative AI Guidelines (2024) | ||||||
| SDAIA AI Ethics Principles (2023) |
| Framework | Inventory | Attribution | Policy log | Redaction proof | Clause map | Signed export |
|---|---|---|---|---|---|---|
| ISO/IEC 42001 | ||||||
| NIST AI RMF |
| Framework | What it expects |
|---|---|
| SAMA Cyber Security Framework | Third-party security, outsourcing and cloud; SAMA approval for material outsourcing |
| NCA Essential Cybersecurity Controls, ECC-2:2024 | Asset management, event logging, external and cloud services; data localisation |
| NCA Cloud Cybersecurity Controls, CCC-2:2024 | Cloud controls updated for data localisation |
| NCA Data Cybersecurity Controls, DCC-1:2022 | Controls across the data lifecycle |
| NCA AI Cybersecurity Guidelines (draft, 2026) | Governance, defence, resilience and third-party controls for generative and agentic AI |
| PDPL, Royal Decree M/19, and its regulations | Processing rules, Transfer Regulations, 72-hour breach notice; fines up to SAR 5 million |
| SDAIA Generative AI Guidelines (2024) | Responsible use of generative AI by government and the public |
| SDAIA AI Ethics Principles (2023) | Accountability, privacy and transparency for AI in use |
| ISO/IEC 42001 | AI management system: inventory, roles, controls and monitoring for AI in use |
| NIST AI RMF | Govern, Map, Measure, Manage |
Answers for Gulf banks, insurers and government entities.
Observe records the app, the account type, the data class and the decision. Prompt text is captured only where a policy requires it, and it stays inside your deployment.
Yes. The endpoint agent finds MCP servers, clients and CLI agents at runtime on Windows, macOS and Linux. The SDK and API gateway cover in-house agents. Tool calls get the same four actions as prompts.
Yes. Kubernetes or VMs with no outbound connection. The app dictionary and detection updates arrive as signed bundles.
Inspection is asynchronous and near real time. The measured p95 will be published after the first pilots, not before.
Not in the on-premise or air-gapped deployment. The backend runs in your data centre in the Kingdom on CPU only. Policy, logs and the Evidence Pack stay inside your perimeter, in line with the localisation provisions in the NCA controls and the PDPL Transfer Regulations.
Per protected user per year, by tier: Browser, Workforce, Enterprise. Quotes are issued at the pilot readout. There is no public price list.
Find your shadow AI on the day you deploy. Reach full shadow AI governance in one week, for one business unit. Delivered by the SOAISEC GCC team with Forcespot, our value-added distributor in the region, and NitronEdge, our value-added solution seller.
The category, the four decisions, what UAE and Saudi regulators expect, and a 12-question vendor checklist.
Read the guideFilter by browser extension, endpoint agent, on premise and price. Compiled from public vendor sites.
Compare the platformsInventory, decisions, agents, exceptions, UAE and Saudi clause maps and a signed export. Printable.
Open the sampleBans push usage to personal accounts. A one-week path from ban to governed use.
Read the articleEight questions the auditor will ask, and the evidence that answers each one.
Read the articleShadow AI found on day 0, an unsanctioned tool moved to sanctioned use, Microsoft 365 Copilot governed, and all four actions live by day 7. Voiceover, captions and transcript.
Watch the video
