cyberAgent — AutoPentest-Qwen2.5-7B-ReAct

• reconnaissance planning
• enumeration workflows
• exploit validation reasoning
• adaptive retry logic
• operational cybersecurity narration
• ReAct-style pentest trajectories

Features

• Cybersecurity reasoning
• Reconnaissance & enumeration planning
• Failure-aware operational thinking
• Multi-turn pentest workflow understanding
• SOC analyst & cyber range simulation
• Human-in-the-loop security copilot

Intended Use

• cybersecurity copilots
• SOC simulation
• CTF assistance
• cyber range environments
• pentest workflow planning
• synthetic trajectory generation
• security education & demos

• fully autonomous offensive operations
• unauthorized security testing
• production exploit automation

Training

Base Model

• Qwen/Qwen2.5-Coder-7B-Instruct

Fine-Tuning

• QLoRA / PEFT
• 4-bit training
• Supervised Fine-Tuning (SFT)

Dataset

• reconnaissance
• web enumeration
• vulnerability assessment
• exploit validation
• retry flows
• operational reasoning
• realistic failures
• ReAct-style conversations

Current Limitations

• Structured JSON/XML outputs may be inconsistent
• Tool-call formatting is not fully deterministic
• Requires human oversight
• Better suited for reasoning than direct execution

Recommended Usage