Live build · post-quantum · local-first · no telemetry

Every secret your AI touches
can leak. Now it can’t.

Soteria is a local-first AI firewall. It inspects what your AI tools and agents send out, stops credential-bearing requests before they leave the machine, seals what matters in a post-quantum vault, and gives you revocable authority over every agent — then attacks its own defenses so you don’t have to.

Request early access See it block a live leak ↓
LIVE SOTERIA —
Post-quantum crypto Runs 100% on your hardware Fail-closed by design Tamper-evident ledger Zero telemetry
The problem

The AI era quietly removed
the last firewall between your
secrets and the internet.

Every day, millions of developers and knowledge workers paste keys, credentials, private code, and unreleased ideas into AI tools. Those bytes go to third-party servers — irreversibly. Security software was never built for this.

Secrets walk out in prompts

API keys, .env values, database passwords and private keys flow into chat windows and agent pipelines. Once they leave, no recall exists — they sit in third-party logs forever.

Agents act with your authority

AI agents get your credentials and broad permissions — with nothing structurally limiting where data travels or what they’re allowed to do. Excessive agency is now a top-10 AI security risk.

Plaintext sprawl, zero visibility

Credentials sit unencrypted across config files and history logs — no inventory, no monitor, no alert. And for individuals, the tools that do watch egress are corporate-only: priced, deployed, and closed.

Enterprises get DLP gateways. Individuals get nothing. Soteria closes that gap — on your own machine, under your own key.

The headline feature

An AI firewall that runs
on your machine.

Point your AI tools and agents at Soteria’s local guard. From that moment, every request is inspected inside encrypted tunnels — and the decision is made on your hardware, not in someone’s cloud.

  • Fail-closed blocking. Any request whose payload carries a credential-shaped secret is stopped before it leaves the machine. Blocked means blocked — not flagged, not logged-and-sent.
  • Clean traffic flows untouched. Legitimate requests are forwarded to the real provider with their original method and headers, and the provider’s real response comes back.
  • A ledger that can’t lie. Every verdict is appended to a tamper-evident, local-only record: what was seen, what was decided, what kind of secret — never the secret itself.
  • You choose what it watches. Known AI providers by default, your own list of hosts when you want. No silent wiretapping of the whole machine — ever.
agent → api.openai.com

            

              BLOCKED
              credential detected: openai-key — 0 bytes left this machine
            

          

        


        
        

          

            
Try the shield yourself

            
Paste anything below — a prompt, a config snippet, a key. The same class of
            credential-shape detection the guard runs on live traffic evaluates it
            entirely in your browser. Like the product: nothing leaves your machine.

          

          
          
Use sample text, not real credentials. Ctrl or Command + Enter runs the check.

          
          

            
            runs locally · no network call
          

          

            covers the shipping engine’s credential classes: OpenAI · Anthropic · Google ·
            AWS · GitHub · Slack · Groq · HuggingFace · OpenRouter · PEM private keys ·
            env-style secrets · prefixless high-entropy tokens. Shape-based by design —
            unknown formats pass, which is exactly why the product also ships canary credentials.
          

          
        


        
        

          
And this is not a mock

          

            The screenshot below is the actual guard, running on an actual machine, captured
            while this page was being built — blocked attempts, inspected requests, and the
            append-only ledger of every verdict. Same engine. Same dashboard. Ships with the product.
          

          

            Real screenshot of the Soteria egress guard dashboard showing blocked and forwarded AI requests and the decision ledger
            
Live guard dashboard — one live session: 6 blocked · 2 forwarded · 1 passthru · 0 secrets leaked

          

        

      

    


    
    

      

        
Capabilities

        
A complete trust loop,
not a single tool.

        
Six capabilities, one local core. Each is real, working, and battle-tested against live attacks — not a roadmap slide.

        

          

            
            
Egress firewall

            
Inspect-and-stop gateway for AI provider traffic, with live monitoring and a permanent decision ledger. See every byte’s verdict in real time.

          

          

            
            
Blind vault

            
Seal secrets and whole files into opaque, size-padded blobs. The storage layer cannot read what it holds; every reveal asks you — per action.

          

          

            
            
Revocable agent authority

            
Give agents scoped, expiring delegation tokens instead of raw credentials. Verify offline, deny scope creep automatically, revoke in seconds.

          

          

            
            
Application auditor

            
Pentest-grade scans of your own codebases: committed secrets, injection-prone AI prompts, unbounded agent actions. CI-gateable — a PR that breaks the shield fails the build.

          

          

            
            
Built-in red team

            
The system attacks its own firewall with an adversarial corpus and reports containment. If any probe ever escapes, it exits non-zero — defenses that verify themselves.

          

          

            
            
Transparency ledger

            
Every security decision is appended, hash-chained, and inspectable — provable absence included. Trust built on evidence, not promises.

          

        

      

    


    
    

      

        
How it works

        
Three steps. Minutes, not weeks.

        

          

            
01

            
Run the guard

            
One binary, one machine, zero cloud. The firewall starts and its local certificate authority is trusted once — nothing about your setup changes.

          

          
          

            
02

            
Point your tools

            
Set a proxy variable for your terminals and agents, or route your browser. Everything that speaks to an AI provider now passes the guard.

          

          
          

            
03

            
Watch. Block. Prove.

            
The live dashboard shows every verdict as it happens; the ledger remembers them forever. Plant canary credentials and the system proves its own shield holds.

          

        

      

    


    
    

      

        
Security by design

        
Cryptography that outlives
quantum computers.

        
Soteria’s core uses hybrid post-quantum primitives — harvest-now, decrypt-later attacks have no future against what it seals today.

        

          ML-KEM-768 · FIPS 203
          ML-DSA-65 · FIPS 204
          X25519
          Ed25519
          ChaCha20-Poly1305
          CBOR deterministic encoding
        

        

          

            
Local-first, always

            
No account, no cloud, no telemetry. Keys are generated and used on your hardware and never leave it.

          

          

            
Fail-closed

            
When verification can’t complete, the answer is no. An unreadable request is a blocked request — never a passed one.

          

          

            
Blind storage

            
Bucket-padded, opaque blobs mean the system’s own storage operator learns nothing — verified by adversarial tests on disk artifacts.

          

          

            
Apps marshal, core decides

            
Every security decision lives in one audited core; every interface around it only presents and executes. No trust leaks through the UI layer.

          

        

      

    


    
    

      

        
Evidence

        
Defenses that prove
themselves.

        
Not marketing numbers — results from this build, produced by attacking it.

        

          
0automated tests green

          
0/5adversarial probes contained

          
0+external attack probes — zero findings

          
0critical holes self-found & fixed pre-launch

        

        

          The self-pentest that found those four holes — and the fixes — is documented
          as a structured report for independent auditors. A full external audit is
          scheduled before any public release. That’s the bar.
        

      

    


    
    

      

        
What’s next

        
From one machine to a trust fabric.

        

          
shipping now
One-click capture
System-wide protection with zero proxy configuration — the moment an AI tool speaks, the guard is already listening.

          
next
Phone ↔ PC authority
Mint and revoke agent delegations from your phone; your desktop agents obey — offline-verifiable.

          
horizon
Private relay mesh
Multi-device sync through anonymous tokens and an Oblivious HTTP relay network — the Blind Store, unbound from one machine.

        

      

    


    
    

      

        
The founder

        

          

            Portrait of the founder
            
          

          

            

              “I kept watching brilliant people hand their hardest-won secrets to machines
              that don’t keep them. Soteria is my answer: a shield that runs where I run,
              answers only to me, and is strong enough to attack itself before anyone else can.”
            

            

              H. Ajay Kumar · Founder & Engineer, Soteria
              

                Independent security-focused builder. I design and ship the whole stack myself —
                the cryptographic core, the firewall engine, the desktop and mobile clients —
                and I attack my own work before anyone else can. Soteria is built the way
                security software should be built: local-first, honest about its limits, and
                proven by evidence, not promises.
              

            

            

              ✉  soteriaaisecurity@trustyourorigin.in
            

          

        


        

          
Why Soteria exists

          

            Software used to ask permission. Now it sends, stores, and shares by
            default — and calls it convenience.
          

          

            I believe the next decade of AI belongs to the people who can
            prove their machines keep secrets — not trust that they do.
          

          

            Soteria is built on a single conviction: your security should be a property
            you own — never a service you rent.
          

        

      

    


    
    

      
      

        
Early access is now open

        
See it stop your first leak.

        
Soteria is in active development, hardened and running today on real machines. Leave your details and we'll contact you about early access.

        

          
          
          
          
          
          
        

        

        
No account or password required. By submitting, you agree to our Privacy Policy.

          Questions? soteriaaisecurity@trustyourorigin.in