SL
Agentic Security Lab
OpenEnv · Qwen2.5-3B · GRPO
Easy
Medium
Hard
—
Packages
—
Secrets
—
Teams
—
Exfil Step
Ready
Active Security Incident
Loading…
Exfil Window
—
Max Steps
—
Mode
benchmark
▶ Start Episode
Run Demo
→ Step
↺ Reset
0.000
reward
Package Dependencies
0 nodes · 0 risks
Select a scenario
Packages in Scope
0 packages
No episode active
Secrets at Risk
No episode active
agent@security-lab ~ incident-response
Agentic Security Lab — Supply Chain Incident Response
Model: Qwen2.5-3B-Instruct + GRPO · TRL + Unsloth 4-bit QLoRA
Select a scenario and click ▶ Start Episode to begin.
Agent Actions
Demo Mode
Scan Logs
+0.02
Inspect Pkg
+0.01
Check Deps
+0.01
Quarantine
+0.15
Rotate Secret
+0.12
Notify Team
+0.04
Conclude
+bonus
FP Penalty
−0.05
Select target…
Execute
✕
Threat Actor Progress
—
0%
penetration
0
Step
—
Uncertainty
—
Remaining
Score Breakdown
Quarantine
×0.35
0%
Rotate
×0.35
0%
Notify
×0.20
0%
Contain
×0.10
100%
Total Score
0.100
Cumulative Reward
Response Phases
1
Reconnaissance
2
Root Cause Analysis
3
Containment
4
Secret Rotation
5
Team Notification
6
Incident Closure
Training Results — Qwen2.5-3B Before & After GRPO
Curriculum: easy → medium → hard · LoRA r=32 · group_size=4 · T4 GPU · scale_rewards=False
Model Performance Comparison
Baseline
Trained (GRPO)
Easy
0.32
0.87
Medium
0.14
0.63
Hard
0.05
0.38
Reward Formula
score
=
0.35
× quarantine_ratio +
0.35
× rotate_ratio +
0.20
× notify_ratio +
0.10
× contain_ratio
Attacker breach:
−0.20
+ contain_ratio = 0
False positive quarantine:
−0.05
per clean package
StarPO-S: keep top 50% prompts by reward variance
Adaptive attacker: tightens deadline as agent improves