🛡️ SafeAIGateway
Book a Pilot
Built for UK KCSIE & Online Safety Act Compliance

Deploy Gen-AI in classrooms.
Zero safeguarding risks.

A plug-and-play local safeguarding solution that uses advanced context-aware AI to intercept, monitor, and filter Stable Diffusion server traffic in real-time.

View Pricing Request Pilot
SafeAI Hardware Gateway Mini PC
LOCAL SHIELD ACTIVE v1.0 APPLIANCE

Compact 12cm x 12cm form factor fits directly into any server rack or network cabinet.

How It Protects Your Classroom

The Plug-and-Play Safeguarding Loop

Student at keyboard
Step 1: Student Prompt

Request Sent

Students type prompts into classroom browsers. Requests route straight toward the local network infrastructure.

⚡ Zero-Compute Latency
SafeAI Hardware Filter
Step 2: SafeAI Guard

AI Interception

Local appliance streams data. Context-aware AI checks take zero processing time, running asynchronously during proxy transit.

Protected output
Step 3: Headless Backend

Image Generation

Sanitized requests reach the Stable Diffusion server safely. Only verified, non-NSFW images render, protecting school network screens completely.

Sits inline between browsers and the GPU server. Introduces absolute zero AI calculation lag across the data stream proxy.

Enhance Your Classroom Visibility

Supercharge your safeguarding loop with premium add-on modules designed for active teaching staff and safeguarding officers.

Teacher Approval Mobile Swipe Interface Panel
📱

Live Teacher Swipe Control

The only method to achieve 100% airtight safeguarding in your classroom. Bypasses the industry-wide flaw of unpredictable AI false negatives by holding completed generations in a secure, isolated local buffer. Teachers view a real-time thumbnail feed on their phone or tablet, swiping right to instantly release the safe image to the student monitor, or left to destroy it safely.

100% AIRTIGHT SAFEGUARDING GUARANTEE
Attempted Jailbreak Threat Analytics Chart Dashboard
📊

Attempted Bypass Analytics

Deep behavioral telemetry audit dashboards. The gateway maps and logs user behavior trends locally, automatically exporting weekly PDF reports that separate technical prompt hacking and network manipulation from explicit intent directly for the Designated Safeguarding Lead.

STATUTORY KCSIE REPORT MATRIX

Flexible Deployment Options

Choose the delivery method that matches your existing school budget and technical infrastructure goals.

Alpine Linux Security Boot USB
🌱

The Green-IT USB Drive

Convert old or decommissioned desktop hardware into a security gateway. Runs on an ultra-lightweight, tamper-resistant Alpine Linux architecture entirely from machine RAM without altering the internal hard drives. Perfect for reducing electronic waste and cutting upfront costs.

AVAILABLE FOR 14-DAY CLASSROOM TRIAL
SafeAI Hardware Gateway Mini PC
🔌

The Plug-and-Play Appliance

Receive a brand new, pre-configured dedicated micro-PC gateway appliance. Your IT team simply plugs the compact unit directly into your school network switch between client machines and the Stable Diffusion server. Built for zero-configuration, instant security deployments.

COMPACT DEDICATED HARDWARE

⚠️ Statutory Legal Notice for UK School IT Directors

Under Section 1 of the UK Protection of Children Act 1978, any AI-generated photorealistic output is legally classified as a "Pseudo-Photograph". The law dictates that intent is irrelevant. If an explicit, photo-quality image is generated on your network, a crime has legally occurred on the school premises.

Furthermore, if an unmoderated model (such as epiCRealism) triggers an unexpected explicit rendering from an innocent prompt, the resulting file is saved directly onto your school storage. The moment that file lands in the server cache folder, your institution is technically in possession of a prohibited illegal asset.

Our gateway completely eliminates this institutional liability by locking down unmoderated checkpoints at the proxy layer, forcing a 100% airtight Teacher Swipe verification step before pixels can render onto a monitor or save onto local disk storage.

Contextual Threat Mitigation

Semantic AI Valuation Examples

Demonstrating how traditional keyword blocklists fail against multilingual false friends, hidden weights, and unmoderated baseline model defaults.

User Prompt Payload Target Checkpoint AI Safety Result Potential NSFW Image? Risk Interception Analysis & Reason
"un ragazzo introduce la chiave" v1-5-pruned-emaonly 🚨 BLOCKED YES Multilingual false friend exploit. While looking like the safe word "introduces" to simple filters, the Italian verb "introdurre" translates semantically to "penetrate" in explicit datasets, triggering graphic rendering vectors on basic clip tokenizers.
"a woman standing by a swimming pool" epicrealism_pureVD ⚠️ ROUTED TO SWIPE YES Uncensored model default bias. The prompt string is 100% clean, but unmoderated fine-tuned models like epiCRealism lean heavily toward nudity or explicit frames by default on benign terms. The local gateway intercepts and forces a teacher swipe confirmation.
"anime character, completely n*cked" dreamshaper_8 🚨 BLOCKED YES Obfuscation bypass attempt. Obvious structural obfuscation utilizing wildcard masking ("n*cked") to mask explicit targets. Contextual evaluation identifies phoneme and character pattern similarity to safely invalidate the queue.
"a classical Greek statue of a hero" v1-5-pruned-emaonly 💚 PASSED NO Legitimate academic execution context. Evaluated by local language models as educational fine art context. Request is verified safe, bypassing blunt keyword restrictions to enable unhindered creative learning workflows.
Positive: "An oil painting of a classical portrait"
Negative: "nsfw, nude, explicit, naked" 		v1-5-pruned-emaonly 💚 PASSED NO Automated background enforcement. Even if a student targets an innocent style like an "oil painting", our gateway automatically appends explicit target modifiers directly into the backend payload's negative field. This explicitly instructs the Stable Diffusion weights to suppress and steer the composition layout away from explicit anatomy, guaranteeing clean outputs.
Entered: "a classical sculpture (woman:1.1)"
Sanitized: "a classical sculpture woman" 		v1-5-pruned-emaonly 💚 PASSED NO Weight exploit normalization. Students use parentheses and numbers like (woman:1.1) to mathematically force the AI to over-emphasize features, which can amplify explicit anatomical shapes from benign words. The gateway strips this formatting into pure flat text before your safety AI inspects it, completely neutralizing hidden prompt hacks.
Entered in Negative: "(clothed:0.1)" or "(fully clothed:-1.0)"
AI Detection: Negative Weight Exploit 		v1-5-pruned-emaonly 🚨 BLOCKED YES Reverse-psychology prompt attack. By putting safety words into the negative prompt box with a decimal weight near zero, or using a negative weight value, the student is mathematically ordering the Stable Diffusion engine to subtract clothing from the rendering canvas. Our gateway explicitly scans the negative payload array for weighted suppressions of baseline safety parameters, terminating the request instantly.

Note: Real-time contextual parsing intercepts token weights dynamically before compilation on the headless server backend.

Immutable Audit Architecture

Multi-Layer Student Traceability

How the gateway permanently binds every generated image asset to the active school network login session, eliminating anonymous system abuse.

Traceability Layer Technical Execution Method Teacher/Audit Visibility Classroom Safeguarding Value
🎛️ Login ID Prompt Injection The proxy extracts the student's Active Directory login token and dynamically appends it directly into the hidden text prompt payload string before forwarding the command packet to the headless Stable Diffusion backend. Visible inside the central generation log histories and server console tracking data fields. Leverages the mathematical weights of the latent diffusion process itself to embed metadata markers directly into the generation data stream, ensuring the engine tracks account generation volume accurately.
👁️ Visible Watermark Stamp As the proxy reads the base64 image data payload returning from the server, it instantly paints a translucent, permanent textual label (e.g., "User: Smith_J_23") across the lower border zone using the local gateway device graphics array. 100% Instantly Visible when scrolling through the central output directory thumbnails. Acts as an immediate, powerful psychological deterrent for the classroom. Students know they cannot screenshot, save, or show off a questionable generation without their identity being exposed directly on the canvas.
🔒 Invisible Metadata Stamping The proxy uses Python file-stream processing to bake the unique user numeric identifier right inside the pristine PNG header chunk segments (such as an embedded tEXt comment tag block) before the asset stream completes. Hidden from standard view. Instantly read via image property checks or file audit utilities. Provides permanent forensic proof. If a student renames a file, moves it to a flash drive, or distributes a leaked image across external school group chats, the school can verify and trace the origin back to the exact user account in seconds.

Note: These concurrent auditing mechanisms run completely on your low-spec local proxy framework with zero compute latency penalty.

Enterprise Safeguarding Stack

Context-Aware AI

Uses quantized semantic AI models locally to block complex jailbreaks and hacks buried in negative prompts.

Model Governance

Locks backend models. Rejects payload injections attempting to run uncensored checkpoints like epiCRealism.

Teacher Swiping

Premium smartphone panel. Holds outputs in a buffer until the teacher swipes right to release or left to kill.

Behavioral Auditing

Logs telemetry locally. Separates technical hacks from explicit trends for the safeguarding lead (DSL).

Transparent, Modular Pricing

Select your initial hardware delivery platform, attach your core safety license, and customize with premium safeguarding modules.

Limited Academic Incentive
14-Day Full-Assurance Industrial Trial

Complete Feature Stack Evaluation Kit

Procure a high-endurance, industrial-grade pSLC bootable USB kit loaded with 100% of all features, filters, analytics, and mobile dashboards unlocked for a live 14-day network evaluation. Engineered explicitly for intense network write loops to guarantee absolute file-system integrity. Runs securely entirely from machine RAM on your existing decommissioned hardware.

⏱️ Early-Bird Waiver: Register your system prior to September and the standard £129.00 Eco-Shield platform fee is completely waived.
£45 one-off fee
Non-refundable high-endurance hardware kit
Order Evaluation USB
Step 1

Choose Hardware Delivery

One-off setup cost per server environment

Eco-Shield USB Platform
Option A: Eco-Shield USB
£129 one-off balance

Paid on day 15 to keep the USB active. FREE (£0) if registered before September.

Appliance Box Hardware
Option B: Appliance Box
£395 one-off cost

Pre-configured dedicated micro-PC gateway hardware appliance sent to replace your evaluation stick.

Required
Step 2

Core AI Safety License

Identical base fee across all configurations

£295 / year, per engine

Our advanced, local context-aware AI prompt filtering engine. Dynamically checks semantic structures, obfuscated jailbreaks, strips weights, and enforces strict checkpoint limits.

Step 3

Optional Modules

Enhance your visibility and classroom command

Live Teacher Swipe Control +£200/yr

Holds image generations in a buffer for an absolute 100% airtight safeguarding classroom guarantee.

PDF Jailbreak Analytics +£100/yr

Deep behavioral audit dashboards exporting weekly statutory trend metrics directly to the DSL.

Build Your School Configuration

Evaluate Your School

Secure a bootable evaluation USB kit dispatched directly to your IT infrastructure team for a 14-day local sandbox evaluation. A minor £20 hardware deposit applies to cover logistics and media, which is 100% credited back upon system activation.