🛡️ VAPT-Env — Live Operations Center
OpenEnv hackathon submission — an environment that teaches a 3-billion-parameter language model to do real security audit reasoning.
Llama 3.2 3B average score 0.075 → 0.482 (6.4× improvement) after GRPO post-training on this env. Real W&B reward curve. Trained adapter on HF Hub. No synthetic data.
🎯 Pick a scenario + agent
📺 Action timeline (live)
What's running
- Live FastAPI env: https://huggingface.co/spaces/Sayuj63/Vapt-env — OpenEnv-compliant, multi-agent v2.0.0
- Trained LoRA adapter: https://huggingface.co/Sayuj63/vapt-env-llama32-3b-grpo — Unsloth 4-bit + LoRA r=16
- W&B training run: https://wandb.ai/sayujpillai63-itm/vapt-env-grpo/runs/ln2jq71s — 112 GRPO steps, real curves
- GitHub: https://github.com/Sayuj63/vapt-env — env + grader + 78 tests + Colab notebook
Hackathon themes covered
- Theme #3.1 World Modeling — partially observable enterprise simulation, 10 security tools, dynamic state
- Theme #1 Multi-Agent —
spawn_subagent/return_to_parentfirst-class actions, budgeted delegation - Theme #2 Long-Horizon — 25/35/45-step audits with sparse rewards + phase tracking
Why hard stays at zero
Hard uses raw HTTP output with honeypots — frontier models (Gemini 2.5 Flash) score ~0.27. Our 3B + harness can't bridge this gap with 28 prompts × 2 epochs of GRPO alone. That's the reasoning gap the env is designed to expose.