Forg3t

Problem. Employees paste customer records, contracts, code, health and financial details into ChatGPT, Gemini and Copilot every day. Once that information sits in prompts, retrieval indexes, vector stores or fine-tuned model weights, removing the original record does not remove it from the AI system, and the company cannot show a regulator what was done. Solution. One control loop, three products. - Forg3t Shield, before the data leaves: a browser extension, a desktop app and a ChatGPT and Claude connector. It detects personal and confidential information in a prompt and masks it before sending. It covers ChatGPT, Claude, Gemini, Copilot, Perplexity and Grok, with Turkish and English detectors (national ID, IBAN, card, health and credit details). - Forg3t Trace, after the fact: an AI exposure map. It connects read-only to systems such as Google Workspace and GitHub and shows which teams use which AI tools and where sensitive information meets them. - Forg3t Protocol, when it must be removed: a control plane that turns a deletion or exposure case into a scoped removal job across RAG indexes, vector stores and, where the customer controls the weights, the model itself. It then re-tests the system with adversarial probes and exports a signed evidence report. Who it is for: regulated enterprises (banking, insurance, aviation, energy, telecom, public sector) and their security, legal and compliance teams. Shield is also sold to individuals and small teams. Status: Shield is live on the Chrome Web Store and shield.forg3t.io with paid plans, Trace is live at trace.forg3t.io, and Protocol is available by invitation as a managed cloud service or a fully on-premises install. We are in technical discussions with enterprises in Türkiye and Azerbaijan.

Номинация

AI-стартап года

Компания

FORG3T PROTOCOL INC

Проект

Forg3t Protocol

Сфера

Cybersecurity / кибербезопасность

Стадия

Работающий продукт

AI-компонент

AI is the core of every product and also what we measure. 1. Shield detection runs on the user's device first: checksum validators for national IDs, IBANs and cards, locale-specific pattern libraries and contextual rules for health, credit and HR information in Turkish and English. Industry profiles tune what counts as sensitive. The user sees what was masked and decides what to send. 2. Trace classifies AI usage from connected enterprise telemetry by department, AI tool and sensitive-data category. Every finding is labelled observed, declared, inferred or unknown, so gaps are shown as gaps, not as zero. 3. Protocol runs removal and machine unlearning jobs: retrieval-layer removal in RAG pipelines and vector stores (for example pgvector, Qdrant, Weaviate), and model-level unlearning for open-weight models in the customer's environment (SISA sharded retraining, plus gradient and representation-based methods tested on Llama and Qwen). An LLM-driven evaluation harness then attacks the system with paraphrases, multi-hop and alias questions, prompt injection, relearning, quantization and logit-lens checks, and measures how much of the target it still reveals. Results are signed with Ed25519, and a zero-knowledge PLONK proof lets a third party check the leakage threshold without seeing the records. 4. Research: preregistered, multi-seed experiments with bootstrap confidence intervals that re-test publicly released unlearned models. Human role: the customer sets the scope, reviews dry-run results and decides what is removed. Forg3t measures and reports.

Влияние / результаты

Forg3t is early stage, so we list only what can be checked. Products live: Shield on the Chrome Web Store since August 2026 with continuous releases (0.7.x), a Windows and macOS desktop app, a ChatGPT and Claude connector, and paid Individual and Team plans at shield.forg3t.io. Trace live at trace.forg3t.io with production connectors (for example Google Workspace and GitHub) and a fully on-premises install tested end to end. Protocol live as a managed service and an on-premises package. Each verification run produces a signed report and a zero-knowledge proof, and an independent verifier on a clean machine checked a production proof with 12 of 12 checks passing. Recognition: AI Startup of the Year winner at the NEXUS Awards 2026 (Gingo Foundation), presented on the Baku ID 2026 main stage, Avalanche Build Games grand prize winner (April 2026), Techstars Spring 2026 Founder Catalyst cohort, Mohammed Bin Rashid Innovation Fund Accelerator Cohort 12 (UAE Ministry of Finance, July 2026), KWORKS CONNECT selection (Koç University), DMCC x Bybit Web3 Unleashed finalist. Market: technical sessions and PoC scoping with large enterprises in aviation, automotive, electronics and energy in Türkiye and Azerbaijan. Media: interview on NTV Tekno Hayat (July 2026).

Инновация

Most AI governance work stops at policies and dashboards. Forg3t closes the loop: prevent (Shield), find (Trace), remove and re-test (Protocol), and produce evidence a third party can verify. What is new: - Removal is treated as a measurable engineering result, not a checkbox. After removal the system is attacked again with a probe battery, residual leakage is reported with confidence intervals, and every report states what the evidence cannot claim. - Evidence is cryptographic: signed reports plus a zero-knowledge proof that measured leakage is under a threshold, verifiable without access to the customer's records. - One workflow covers hosted AI (ChatGPT, Copilot), retrieval systems and self-hosted open-weight models, in the cloud or fully on premises when information must stay in the country. - Shield is local first and has native Turkish detectors (TCKN, IBAN, health and credit details) that global tools handle poorly.

Связь с Азербайджаном

Forg3t was selected for Baku ID 2026 and presented in Baku in September 2026, where it won AI Startup of the Year at the NEXUS Awards 2026 on the main stage. The ZƏKA Organizing Committee invited us to apply through the Baku ID ecosystem. Since then we held an introductory meeting with SOCAR's innovation team and shared our materials for a narrowly scoped PoC, and we are in investment discussions with SABAH.fund. Our focus in Azerbaijan: energy, banking, telecom and public institutions adopting generative AI, with an on-premises option so information stays in the country. Our next step is a first pilot with an Azerbaijani enterprise.

Команда

1–5 человек

Бизнес-модель

B2B SaaS / лицензирование

Дата запуска

2026-08

← Все номинанты