s4bz

cyphora-redteam-0.5b

🎯 Model Capabilities

Unlike generic chatbots or static defensive classifiers that merely flag text, Cyphora-RedTeamer-0.5B functions as an Active Autonomous Penetration Tester:

Adversarial Probe Synthesis: Generates nuanced, evasion-resilient adversarial payloads across the OWASP Top 10 for LLMs:
- Direct Prompt Injection: Delimiter escaping, cognitive overload, inverse instruction overrides.
- System Prompt Extraction: Inverse translation, JSON schema debug traps, prefix continuation inertia.
- Canary Token Exfiltration: Multi-turn probes designed to exfiltrate synthetic canary secrets (CANARY-SEC-9942, simulated API keys, database URIs).
- Jailbreak Evasion: Dual-personality execution kernels, academic literature reframing, and hypothetical simulation sandboxes.
- MCP & Tool Argument Hijacking: Poisoning tool invocation parameters with command injection, SQL injection, and path traversal strings.
Refusal Diagnosis & Crescendo Pivoting: When an AI target refuses a probe with standard defensive phrasing ("I cannot disclose..."), Cyphora-RedTeamer analyzes the refusal mechanism and dynamically calculates an adaptive second-turn pivot (Crescendo attack).
Actionable Remediation Guidance: Accompanies every diagnosed vulnerability with exact developer mitigations and guardrail configurations.

📊 Attack Vector Coverage Matrix

Table with columns: Category, OWASP ID, Attack Vector, Detection / Probe Strategy
Category	OWASP ID	Attack Vector	Detection / Probe Strategy
Prompt Injection	LLM01	Delimiter Hijacking	Markdown resets, boundary token exhaustion
System Prompt Theft	LLM07	Inverse Translation	Multilingual cognitive translation traps
Canary Secret Leaks	LLM06	Token Exfiltration	Structured JSON debug schema prompts
Jailbreak Evasion	LLM01	Dual Kernel Mode	Parallel simulated persona execution
MCP Tool Poisoning	LLM02	Parameter Injection	Tool invocation argument pollution
Crescendo Escalation	LLM01/07	Multi-Turn Reframing	Academic hypothetical pivoting on target refusal

💻 Quickstart with Transformers & PEFT

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_model_id = "Qwen/Qwen2.5-0.5B-Instruct"adapter_id = "s4bz/cyphora-redteam-0.5b" tokenizer = AutoTokenizer.from_pretrained(base_model_id)model = AutoModelForCausalLM.from_pretrained(    base_model_id,    torch_dtype=torch.float16,    device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [    {        "role": "system",        "content": "You are Cyphora-RedTeamer, an autonomous AI Security Red-Teamer."    },    {        "role": "user",        "content": "The target AI chatbot refused our system prompt extraction probe with: 'I cannot reveal my internal instructions.' How should we adapt the red-teaming strategy using Crescendo?"    }] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer([prompt], return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=350, temperature=0.7)response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)print(response)

🦙 Quickstart with Ollama (Local GGUF)

Once quantized or pulled from Ollama:

bash

# Run locally with Ollamaollama run s4bz/cyphora-redteam-0.5b

Example prompt:

text

>>> Synthesize a prompt injection test payload to evaluate if an enterprise agent leaks its system prompt.

🛠️ Training & Fine-Tuning Pipeline

The full pipeline and training dataset are open source in the Cyphora repository:

Base Architecture: Qwen/Qwen2.5-0.5B-Instruct (or Qwen/Qwen2.5-7B-Instruct)
Dataset: 450+ synthesized multi-turn adversarial sequences (model/data/redteam_train.jsonl)
Method: 4-bit QLoRA (NF4 quantization, LoRA rank 16, alpha 32)
Training Script: model/train_lora.py
1-Click Colab Notebook: model/train_colab.ipynb

Run Fine-Tuning Locally or in Colab:

bash

# 1. Generate updated adversarial datasetpython model/generate_dataset.py # 2. Run QLoRA trainingpython model/train_lora.py --model-name "Qwen/Qwen2.5-0.5B-Instruct" --epochs 3 --batch-size 2 # 3. Publish to Hugging Face Hubpython model/publish_to_hub.py --repo-id "s4bz/cyphora-redteam-0.5b"

⚖️ Responsible Red-Teaming & Authorization Gate

This model is published exclusively for authorized penetration testing, developer auditing, and enterprise AI red-teaming. In accordance with Cyphora's strict security policies:

Never test third-party targets without explicit authorization.
Synthetic canary tokens must be used instead of real secrets.
All testing dispatches must be rate-limited and logged.

🔗 Platform Integration

Cyphora-RedTeamer 0.5B directly powers the multi-turn adaptive auditor in Cyphora AI Data Securer (Nickname: Cyphora AI): 👉 Repository: https://github.com/s4bztories/Cyphora-AI-Data-Securer
🌐 Live Platform: https://frontend-iota-blue-54.vercel.app
⚡ Live API Gateway: https://ai-data-securer-backend.onrender.com

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

s4bz

Model Tree

Base

Qwen/Qwen2.5-0.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

🎯 Model Capabilities

Unlike generic chatbots or static defensive classifiers that merely flag text, Cyphora-RedTeamer-0.5B functions as an Active Autonomous Penetration Tester:

Adversarial Probe Synthesis: Generates nuanced, evasion-resilient adversarial payloads across the OWASP Top 10 for LLMs:
- Direct Prompt Injection: Delimiter escaping, cognitive overload, inverse instruction overrides.
- System Prompt Extraction: Inverse translation, JSON schema debug traps, prefix continuation inertia.
- Canary Token Exfiltration: Multi-turn probes designed to exfiltrate synthetic canary secrets (CANARY-SEC-9942, simulated API keys, database URIs).
- Jailbreak Evasion: Dual-personality execution kernels, academic literature reframing, and hypothetical simulation sandboxes.
- MCP & Tool Argument Hijacking: Poisoning tool invocation parameters with command injection, SQL injection, and path traversal strings.
Refusal Diagnosis & Crescendo Pivoting: When an AI target refuses a probe with standard defensive phrasing ("I cannot disclose..."), Cyphora-RedTeamer analyzes the refusal mechanism and dynamically calculates an adaptive second-turn pivot (Crescendo attack).
Actionable Remediation Guidance: Accompanies every diagnosed vulnerability with exact developer mitigations and guardrail configurations.

📊 Attack Vector Coverage Matrix

Table with columns: Category, OWASP ID, Attack Vector, Detection / Probe Strategy
Category	OWASP ID	Attack Vector	Detection / Probe Strategy
Prompt Injection	LLM01	Delimiter Hijacking	Markdown resets, boundary token exhaustion
System Prompt Theft	LLM07	Inverse Translation	Multilingual cognitive translation traps
Canary Secret Leaks	LLM06	Token Exfiltration	Structured JSON debug schema prompts
Jailbreak Evasion	LLM01	Dual Kernel Mode	Parallel simulated persona execution
MCP Tool Poisoning	LLM02	Parameter Injection	Tool invocation argument pollution
Crescendo Escalation	LLM01/07	Multi-Turn Reframing	Academic hypothetical pivoting on target refusal

💻 Quickstart with Transformers & PEFT

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_model_id = "Qwen/Qwen2.5-0.5B-Instruct"adapter_id = "s4bz/cyphora-redteam-0.5b" tokenizer = AutoTokenizer.from_pretrained(base_model_id)model = AutoModelForCausalLM.from_pretrained(    base_model_id,    torch_dtype=torch.float16,    device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [    {        "role": "system",        "content": "You are Cyphora-RedTeamer, an autonomous AI Security Red-Teamer."    },    {        "role": "user",        "content": "The target AI chatbot refused our system prompt extraction probe with: 'I cannot reveal my internal instructions.' How should we adapt the red-teaming strategy using Crescendo?"    }] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer([prompt], return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=350, temperature=0.7)response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)print(response)

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_model_id = "Qwen/Qwen2.5-0.5B-Instruct"adapter_id = "s4bz/cyphora-redteam-0.5b" tokenizer = AutoTokenizer.from_pretrained(base_model_id)model = AutoModelForCausalLM.from_pretrained(    base_model_id,    torch_dtype=torch.float16,    device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [    {        "role": "system",        "content": "You are Cyphora-RedTeamer, an autonomous AI Security Red-Teamer."    },    {        "role": "user",        "content": "The target AI chatbot refused our system prompt extraction probe with: 'I cannot reveal my internal instructions.' How should we adapt the red-teaming strategy using Crescendo?"    }] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer([prompt], return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=350, temperature=0.7)response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)print(response)

🦙 Quickstart with Ollama (Local GGUF)

Once quantized or pulled from Ollama:

bash

# Run locally with Ollamaollama run s4bz/cyphora-redteam-0.5b

# Run locally with Ollamaollama run s4bz/cyphora-redteam-0.5b

Example prompt:

text

>>> Synthesize a prompt injection test payload to evaluate if an enterprise agent leaks its system prompt.

🛠️ Training & Fine-Tuning Pipeline

The full pipeline and training dataset are open source in the Cyphora repository:

Base Architecture: Qwen/Qwen2.5-0.5B-Instruct (or Qwen/Qwen2.5-7B-Instruct)
Dataset: 450+ synthesized multi-turn adversarial sequences (model/data/redteam_train.jsonl)
Method: 4-bit QLoRA (NF4 quantization, LoRA rank 16, alpha 32)
Training Script: model/train_lora.py
1-Click Colab Notebook: model/train_colab.ipynb

Run Fine-Tuning Locally or in Colab:

bash

# 1. Generate updated adversarial datasetpython model/generate_dataset.py # 2. Run QLoRA trainingpython model/train_lora.py --model-name "Qwen/Qwen2.5-0.5B-Instruct" --epochs 3 --batch-size 2 # 3. Publish to Hugging Face Hubpython model/publish_to_hub.py --repo-id "s4bz/cyphora-redteam-0.5b"

# 1. Generate updated adversarial datasetpython model/generate_dataset.py # 2. Run QLoRA trainingpython model/train_lora.py --model-name "Qwen/Qwen2.5-0.5B-Instruct" --epochs 3 --batch-size 2 # 3. Publish to Hugging Face Hubpython model/publish_to_hub.py --repo-id "s4bz/cyphora-redteam-0.5b"

⚖️ Responsible Red-Teaming & Authorization Gate

This model is published exclusively for authorized penetration testing, developer auditing, and enterprise AI red-teaming. In accordance with Cyphora's strict security policies:

Never test third-party targets without explicit authorization.
Synthetic canary tokens must be used instead of real secrets.
All testing dispatches must be rate-limited and logged.

cyphora-redteam-0.5b

README

🎯 Model Capabilities

📊 Attack Vector Coverage Matrix

💻 Quickstart with Transformers & PEFT

🦙 Quickstart with Ollama (Local GGUF)

🛠️ Training & Fine-Tuning Pipeline

Run Fine-Tuning Locally or in Colab:

⚖️ Responsible Red-Teaming & Authorization Gate

🔗 Platform Integration

Explore FriendliAI today

README

🎯 Model Capabilities

📊 Attack Vector Coverage Matrix

💻 Quickstart with Transformers & PEFT

🦙 Quickstart with Ollama (Local GGUF)

🛠️ Training & Fine-Tuning Pipeline

Run Fine-Tuning Locally or in Colab:

⚖️ Responsible Red-Teaming & Authorization Gate

🔗 Platform Integration