The Sovereign AI Privacy Gateway. Use the world's strongest AI models — with the sensitive values CoreShield detects replaced by tokens before anything leaves your environment.
Find sensitive values — names, IDs, IBANs, medical terms.
Cloak
Replace them with anonymous tokens; originals sealed in an AES-256-GCM vault.
Send
Detected sensitive values are replaced with tokens before sending; values the detection misses — notably some names — can reach the provider as written.
Restore
Rehydrate the AI response locally — a hallucination guard checks the restored values.
Software work registered with the Saudi Authority for Intellectual Property (SAIP) — Registration No. 26-12-122599879— 12 Nov 2026
INTERACTIVE DEMO
See the four verbs in action
Step through Detect, Cloak, Send, and Restore on a sample document. This is a scripted simulation — nothing leaves your browser and no AI model is called.
Demo · simulated
CoreShield scans the document and finds 7 sensitive values — names, IDs, IBANs, amounts, and dates.
Your document
This Supply Agreement is entered into between Najd Petrochemicals Company, registered in Riyadh, Saudi Arabia, and Helvetia Trading AG. The buyer shall remit USD 4,500,000 to account SA44 2000 0001 2345 6789 1234 no later than 15 March 2027. The agreement is authorized by Faisal Al-Dossari, Chief Procurement Officer.
Every accuracy figure in this section comes from one independent synthetic test set, measured on the detector configuration in service now. The set was frozen with SHA-256 before the detector changes it measures and scored exactly once. A figure is the share of planted values detected, shown with its 95% confidence interval: measured, not a guarantee.
Structured identifiers
National ID, Iqama, commercial registration, IBAN and email, combined.
99.0%Measured protection rate
95% CI 98.8–99.2
National ID100.0%precision 98.5%
Iqama100.0%precision 98.5%
Commercial registration93.1%precision 99.9%
IBAN100.0%precision 100.0%
Email100.0%precision 100.0%
National ID and Iqama share one precision figure: the detector flags both as an ID number. Employee numbers are not part of this group; they are shown on the "How we measure" page.
Free-form values are harder to detect than structured identifiers. Names and phone numbers are detected by rules; organisations by rules together with a trained model.
Names95.3%
detected · 95% CI 94.8–95.7 · precision 94.8%
Organisations85.3%
detected · 95% CI 84.2–86.3 · precision 98.0% · rules + trained model
Phone numbers96.4%
detected · 95% CI 95.9–96.9 · precision 99.7%
From the evaluation workspaces only (synthetic data). Refreshed hourly.
Synthetic documents processed
2,294
Values cloaked
119,786
Audit chain verified
257,382 events, intact
Last updated
16 Nov 2026, 21:51
The detector in service is svc-detector v1.34. bench-files-v7 is the only set measured on it. Older sets were measured on earlier versions; they are on the "How we measure" page as history and are not mixed into these figures.
The accuracy figures come from running the detector code on synthetic test documents, outside the live service; the live counters come from evaluation workspaces. No customer data is included.
Frontier models (GPT, Claude, Gemini) drive real productivity. Your competitors are using them today.
The Data Risk
Every prompt sent to an external LLM carries a copy of your sensitive data outside your jurisdiction — outside PDPL, outside NCA.
The CoreShield Solution
Detect. Cloak. Send. Restore. Detected sensitive values are tokenized before they leave your environment, reducing exposure while you keep the gains the AI provides.
Who CoreShield Is For
The same four verbs — Detect, Cloak, Send, Restore — whether you are one person with one contract or a ministry with a million records.
Individuals
Review your own contracts, reports, and statements with frontier AI. Detected names, national IDs, IBANs, and phone numbers are cloaked before anything is sent.
Businesses
Thirteen sector policy packs, team roles, and human review — from a solo consultant to an enterprise, with mutations, AI calls, and vault reads recorded in a hash-chained audit log.
Government
Deployable inside your own data center, bilingual Arabic and English by default, and built around PDPL and NCA requirements.
Built for Every Sensitive Sector
Thirteen sector packs with tuned entity detection, cloaking rules, and review workflows.