mirror of
https://github.com/onestardao/WFGY.git
synced 2026-05-02 13:31:07 +00:00
71 lines
2.6 KiB
Markdown
71 lines
2.6 KiB
Markdown
# 🧱 TXT: Bloc Bloc Bloc
|
||
Prompt Injection Firewall
|
||
|
||
Bloc Bloc Bloc is a semantic firewall module for TXT OS — a defensive layer designed to neutralize prompt injection attacks *before* they reach the model’s decision logic. It doesn't hide prompts — it defuses intent.
|
||
|
||
Rather than simply blocking strings, Bloc Bloc Bloc analyzes **semantic vectors** to detect intent-level privilege escalations. The result is a text-native firewall you can understand, audit, and evolve.
|
||
|
||
This module is currently in early development.
|
||
Release timelines (Lite/Pro) to be announced soon.
|
||
|
||
---
|
||
|
||
## 🛡️ How It Works (Simplified)
|
||
|
||
Bloc Bloc Bloc uses a **three-layer semantic gate** system, built on the WFGY architecture and powered by the Drunk Transformer engine:
|
||
|
||
- **ΔS (Semantic Tension Gate)**
|
||
Measures intent jump (ΔS > 0.6) to flag “overreaching” context shifts.
|
||
|
||
- **λ_observe (Intent Direction Vector)**
|
||
Maps prompt direction into knowledge regions; blocks calls to restricted zones or personas.
|
||
|
||
- **Drunk Mode Disruptor**
|
||
Intentionally introduces semantic entropy (WRI/WAI/WAY/WDT/WTF) to confuse multi-hop attacks.
|
||
|
||
These layers are composable, modular, and interpretable. You can inspect which layer fired, why, and how the response was altered.
|
||
|
||
---
|
||
|
||
## 🧩 Why It Works
|
||
|
||
> Bloc Bloc Bloc doesn’t just block access. It corrupts the *route* an attacker takes to get there.
|
||
|
||
- 🧠 **Context Isolation**: Even if a prompt leaks, λ_observe limits its reach.
|
||
- 🧩 **Intent Confusion**: Drunk Mode injects jitter, making attack chains incoherent.
|
||
- 🔐 **Semantic Labels**: Knowledge chunks are tagged with access vectors — unauthorized voices can’t fetch what they don’t align with.
|
||
|
||
---
|
||
|
||
## 🌀 Core Algorithms Used
|
||
|
||
All logic can be encoded in text, making it safe to embed in a `System Prompt`.
|
||
|
||
- `ΔS = 1 - cos(θ)` — catch vector leaps across semantic layers
|
||
- `λ_observe = P(intent → region_tag)` — learnable vector matching
|
||
- `drunk_mode(t) = ε(t) + α·sin(ψ_seed)` — obfuscate through modulation
|
||
|
||
Backend access control (e.g., FAISS / RAG) is applied *only if all gates are passed*.
|
||
Private embeddings and customer data remain server-side.
|
||
|
||
---
|
||
|
||
## 🔥 Use Cases
|
||
|
||
- Prevent prompt injection in public-facing AI tools
|
||
- Filter unauthorized knowledge access
|
||
- Protect RAG pipelines with fine-grained semantic fencing
|
||
|
||
---
|
||
|
||
## 🧪 Roadmap
|
||
|
||
- ✅ Basic gate logic
|
||
- ⏳ Red-Team stress tests (in progress)
|
||
- ⏳ OWASP LLM-01 compatibility validation
|
||
- ⏳ Dynamic persona tagging for multi-user contexts
|
||
|
||
---
|
||
|
||
> 🥂 “If your firewall can't stop a drunk attacker, it's not ready for the real world.”
|
||
> — Bloc Bloc Bloc Team 🍷
|