Stops sensitive data leaking into ChatGPT, Claude, Gemini, Copilot, and Perplexity — scanned in the browser, before it ever sends.
People paste things into AI chatbots without thinking twice as an AWS key while debugging, a customer's email while drafting a support reply, a chunk of proprietary code while asking for a refactor. None of it is malicious. All of it is a problem.
ShadowAI sits between the employee and the AI tool, reads the prompt in real time, and decides whether it's safe to send entirely inside the browser, before a single character reaches an external server.
No blind trust. No silent blocking either every decision comes with a reason attached.
flowchart TD
A["Employee"] --> B["Browser Extension"]
B -->|intercepts before send| C["FastAPI Backend"]
C --> D["Detection Engine"]
D --> E["Risk Scoring Engine"]
E --> F{"Decision Engine"}
F -->|0-30| G["Allow"]
F -->|31-70| H["Warn"]
F -->|71-100| I["Block"]
F -->|when possible| J["Auto-Redact"]
G --> K["Audit Logs"]
H --> K
I --> K
J --> K
K --> L["Admin Dashboard"]
style A fill:#1E293B,stroke:#64748B,color:#fff
style B fill:#0F172A,stroke:#3B82F6,color:#fff
style C fill:#0F172A,stroke:#3B82F6,color:#fff
style D fill:#0F172A,stroke:#8B5CF6,color:#fff
style E fill:#0F172A,stroke:#8B5CF6,color:#fff
style F fill:#1E293B,stroke:#F59E0B,color:#fff
style G fill:#0F172A,stroke:#22C55E,color:#fff
style H fill:#0F172A,stroke:#F59E0B,color:#fff
style I fill:#0F172A,stroke:#EF4444,color:#fff
style J fill:#0F172A,stroke:#06B6D4,color:#fff
style K fill:#1E293B,stroke:#64748B,color:#fff
style L fill:#1E293B,stroke:#64748B,color:#fff
The whole round trip intercept, scan, score, decide happens before the prompt is ever allowed to leave the machine.
ShadowAI doesn't rely on one model doing all the work. Three detection modules run in parallel, each tuned for a different kind of leak.
flowchart LR
P["Incoming Prompt"] --> S["Secret Detection"]
P --> PII["PII Detection"]
P --> SC["Source Code Detection"]
S --> R["Explainable Risk Score"]
PII --> R
SC --> R
style P fill:#1E293B,stroke:#64748B,color:#fff
style S fill:#0F172A,stroke:#EF4444,color:#fff
style PII fill:#0F172A,stroke:#3B82F6,color:#fff
style SC fill:#0F172A,stroke:#22C55E,color:#fff
style R fill:#1E293B,stroke:#F59E0B,color:#fff
| Module | Catches | Powered by |
|---|---|---|
 |
AWS keys, GitHub tokens, JWTs, passwords, API keys, OAuth tokens, SSH keys | Regex + pattern matching |
 |
Emails, phone numbers, credit cards, Aadhaar, PAN, passport numbers, customer names | Microsoft Presidio, spaCy NER, regex |
 |
Python, Java, JavaScript, C++, Go, SQL | Keyword, import, class/function, and file-structure heuristics |
Running these in parallel instead of stacking them keeps accuracy high without drowning users in false positives.
Most tools in this space hand you a number and call it a day. ShadowAI shows its work.
Risk Score: 87
AWS Access Key +40
Customer Email +20
Source Code +15
Public AI Platform +12
Every risk score is fully explainable, with each point directly linked to the specific sensitive content or behavior detected by the engine—ensuring complete transparency, not blind trust.
| Score | Action | Behavior |
|---|---|---|
0–30 |
 |
Prompt goes through, no friction |
31–70 |
 |
User sees the risk and chooses: cancel, send anyway, or auto-redact — either way, it's logged |
71–100 |
 |
Prompt is stopped per policy, with a clear explanation shown on screen |
Blocking someone outright kills their momentum. Where it's safe to do so, ShadowAI just strips the sensitive part and lets the person keep working.
|
Before |
After |
flowchart TD
A["Employee"] --> B["Browser Extension"]
B --> C["FastAPI Backend"]
C --> D["Detection Engine"]
subgraph DET[" "]
D1["Secret Detection"]
D2["PII Detection"]
D3["Source Code Detection"]
end
D --> D1
D --> D2
D --> D3
D1 --> E["Explainable Risk Scoring"]
D2 --> E
D3 --> E
E --> F["Decision Engine"]
subgraph DEC[" "]
F1["Allow"]
F2["Warn"]
F3["Block"]
F4["Auto-Redact"]
end
F --> F1
F --> F2
F --> F3
F --> F4
F1 --> G["Audit Logs"]
F2 --> G
F3 --> G
F4 --> G
G --> H["Admin Dashboard"]
style A fill:#1E293B,stroke:#64748B,color:#fff
style B fill:#0F172A,stroke:#3B82F6,color:#fff
style C fill:#0F172A,stroke:#3B82F6,color:#fff
style D fill:#0F172A,stroke:#8B5CF6,color:#fff
style D1 fill:#0F172A,stroke:#EF4444,color:#fff
style D2 fill:#0F172A,stroke:#3B82F6,color:#fff
style D3 fill:#0F172A,stroke:#22C55E,color:#fff
style E fill:#1E293B,stroke:#F59E0B,color:#fff
style F fill:#1E293B,stroke:#F59E0B,color:#fff
style F1 fill:#0F172A,stroke:#22C55E,color:#fff
style F2 fill:#0F172A,stroke:#F59E0B,color:#fff
style F3 fill:#0F172A,stroke:#EF4444,color:#fff
style F4 fill:#0F172A,stroke:#06B6D4,color:#fff
style G fill:#1E293B,stroke:#64748B,color:#fff
style H fill:#1E293B,stroke:#64748B,color:#fff
|
The first checkpoint. Before "send" actually does anything, the extension grabs the prompt, holds it, and waits for a verdict. Captured per request:
|
Built on FastAPI and Python. Handles auth, runs the full detection pipeline, applies org-level policy, computes the risk score, logs the event, and returns a decision. |
Every scan is logged user, timestamp, destination, score, trigger, decision, override, and policy applied. Turns raw logs into prompt history, risk analytics, and department-level trends. |
Built to keep people productive with AI — without handing over the keys.