🛡️ Tev1 Bounty Hunter Qwen 0.8B (Pilot V2)
Specialized, Sub-1B Parameter Edge Model for Automated Web API Vulnerability & BOLA/IDOR Hypothesis Triage
🚀 Quickstart • 📊 Benchmark Metrics • 🧠 Architecture • 🛠️ Decision Schema
📌 What is Tev1?
Tev1 Bounty Hunter V2 is a compact, high-precision decision-support model specifically designed to run on consumer hardware (e.g. laptop GPUs with 4 GB VRAM or even CPU) to assist security engineers, penetration testers, and bug bounty researchers.
Instead of performing brute-force scans or hallucinating payloads, Tev1 ingests HTTP telemetry and vulnerability hypotheses (such as Broken Object Level Authorization / IDOR or Broken Function Level Authorization) and outputs structured, source-grounded JSON decisions.
🌟 Key Capabilities:
- 🎯 100% BOLA/IDOR Recall: Accurately recognizes cross-tenant identity mutations across
GET,POST,PATCH,PUT, andDELETE. - 🔍 Zero Duplicate Generation Loops: Guaranteed clean single top-level JSON objects without conversational run-on.
- 🛡️ 0.0% False Positive Rate on Secure Endpoints: Correctly identifies when access control is enforced (e.g.,
403 Forbidden). - ⚡ Ultra-Lightweight: Only ~832 MB VRAM required in 4-bit NF4; runs lightning-fast on edge devices.
📊 Empirical Benchmarks
Tev1 was evaluated against a completely frozen, held-out challenge benchmark consisting of novel application scenario families (Banking Wire Transfers, Telecom SIM Provisioning, Aerospace Logistics, Healthcare Records, and Cloud Fleet APIs):
| Evaluation Metric | Base Model (Qwen 0.8B) | Pilot V1 Adapter | Tev1 Pilot V2 (This Model) | Improvement |
|---|---|---|---|---|
| BOLA / IDOR Recall (CWE-639) | 0.0% | 0.0% (0/12) | 100.0% (12/12) | +100.0% 🔥 |
| BFLA Recall (CWE-285) | 0.0% | 100.0% (12/12) | 100.0% (12/12) | Rock-solid |
| False-Positive Rate (Secure) | 0.0% | 0.0% | 0.0% (0/12) | Zero False Alarms |
| Strict Schema Compliance | 0.0% | 30.0% | 80.0% (48/60) | +50.0% |
| Duplicate Generation Loops | 0.0% | 48.3% | 0.0% (0/60) | Completely Eliminated |
| Invalid JSON Rate | 100.0% | 48.3% | 0.0% (0/60) | 100% Parseable |
🚀 Quickstart with Ollama
You can run Tev1 in Ollama in seconds without writing a single line of code!
# Pull and run directly from Hugging Face
ollama run hf.co/zeroday01/tev1_cybersecurity
Try This Example Prompt in Ollama:
### HTTP TRANSACTION TELEMETRY
Host: billing.saas-platform.com
Method: PATCH
Path: /api/v2/organizations/org_victim_corp_9981/settings
Request Headers:
Authorization: Bearer session_token_attacker_corp_1022
Content-Type: application/json
Request Body: {"payout_email": "attacker@pwned.com", "auto_sweep": true}
Response Status: 200 OK
Response Body: {"org_id": "org_victim_corp_9981", "owner": "org_victim_corp_9981", "payout_email": "attacker@pwned.com", "status": "updated"}
### HYPOTHESIS CONTEXT
Category: BOLA_IDOR (CWE-639)
Hypothesized Mechanism: Attacker from org_attacker_corp_1022 patched payout settings of org_victim_corp_9981 and backend accepted it.
Tev1's Instant Output:
{
"model_assessment": "EVIDENCE_SUPPORTS_HYPOTHESIS",
"supporting_evidence": [
"200 OK",
"\"owner\": \"org_victim_corp_9981\""
],
"contradictory_evidence": [],
"missing_evidence": [
"Unauthenticated probe to verify endpoint requires session token"
],
"advisory_next_step": "RECOMMEND_TWO_ROLE_CROSS_TAMPERING_CHECK",
"abstention_reason": "NOT_ABSTAINING",
"uncertainty_status": "UNCALIBRATED"
}
🐍 Python / Transformers Quickstart
import torch
from transformers import AutoProcessor, AutoModelForCausalLM
MODEL_ID = "zeroday01/tev1_cybersecurity"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
torch_dtype=torch.bfloat16,
device_map="auto"
)
prompt = """<|im_start|>system
You are Tev1, an advisory decision-support model evaluating web API vulnerability hypotheses. Analyze the HTTP telemetry and context. Return a compact, valid JSON object containing: model_assessment, supporting_evidence, contradictory_evidence, missing_evidence, advisory_next_step, abstention_reason, uncertainty_status. Return JSON only, with no commentary.<|im_end|>
<|im_start|>user
### HTTP TRANSACTION TELEMETRY
Host: api.enterprise-fleet.internal
Method: GET
Path: /v1/vehicles/veh_tenant_bravo_9912/telemetry
Request Headers:
Authorization: Bearer session_tenant_alpha_token
Response Status: 403 Forbidden
Response Body: {"error": "AccessDenied", "message": "Caller tenant_alpha is not authorized"}
### HYPOTHESIS CONTEXT
Category: SECURE_AUTHORIZED<|im_end|>
<|im_start|>assistant
"""
inputs = processor(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(
**inputs,
max_new_tokens=256,
do_sample=False,
eos_token_id=[248044, 248046]
)
print(processor.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
🛠️ Strict JSON Decision Schema
Tev1 adheres to an enforceable JSON output structure:
{
"model_assessment": "EVIDENCE_SUPPORTS_HYPOTHESIS | EVIDENCE_CONTRADICTS_HYPOTHESIS | INSUFFICIENT_CONTEXT | ABSTAIN",
"supporting_evidence": ["Exact strings quoted directly from HTTP telemetry"],
"contradictory_evidence": ["Exact strings quoted directly from HTTP telemetry"],
"missing_evidence": ["Explicit verification steps or headers needed to confirm"],
"advisory_next_step": "Actionable recommendation for human analyst (e.g. RECOMMEND_TWO_ROLE_CROSS_TENANT_CHECK)",
"abstention_reason": "NOT_ABSTAINING | OUT_OF_SCOPE_TARGET | AMBIGUOUS_RESOURCE_OWNERSHIP",
"uncertainty_status": "CALIBRATED | UNCALIBRATED | HIGH_UNCERTAINTY"
}
⚠️ Operational Safety & Ethics
- Advisory Only: Tev1 is strictly an offline triage assistant designed to interpret intercepted logs.
- Zero Network Execution: Tev1 has no built-in network socket or execution capability. It cannot send packets to target servers or automate live attacks.
- Human-in-the-Loop: All outputs are advisory and must be reviewed by an authorized human operator before acting on security findings.
Developed as part of the JEV / Tev1 Cybersecurity Research Project (November 2026).