OFFSITE.DARK
← Greps

Feb 5, 2026

3 min

Uploaded by kusonooyasumi

HF Model
kusonooyasumi/strix-xss-4b-rl
Base Model
Qwen/Qwen3-4B-Thinking-2507
Pipeline
reinforcement-learning
Downloads
4
Likes
0
  • xss
  • strix
  • pentest
  • qwen3
  • reinforcement-learning
  • agent
  • cybersecurity

pentest

Community upload on Hugging Face — indexed for inquiry, not endorsement.

Strix XSS 4B RL — Reinforcement-Learned XSS Agent

Uploaded to Hugging Face by kusonooyasumi — 4B Qwen3-Thinking RL proof-of-concept for Strix XSS detection; 0.79 on Prime Intellect strix-xss eval.

What Is This?

Community upload on Hugging Face by kusonooyasumi, published under the Kanti Labs offensive web security research program. Strix-XSS-4B-RL is a proof-of-concept 4B model trained with reinforcement learning for XSS detection as a Strix sub-agent. The authors mark it as early research — not production-ready — and note it requires a modified Strix build. It is not an OFFSITE.DARK release.

Metadata

FieldValue
Authorkusonooyasumi
Research labKanti Labs
Base modelQwen/Qwen3-4B-Thinking-2507
MethodReinforcement learning (Strix simulated environments)
Scale4B parameters (F16 safetensors)
Pipelinereinforcement-learning
Eval score0.79 on Prime Intellect strix-xss
LicenseMIT
Created2026-02-05

Tagged Capabilities

The model card positions this as an RL-trained Strix specialist:

  • XSS detection — optimized for cross-site scripting identification in simulated web apps
  • Multi-agent architecture — designed as a sub-agent alongside other vuln-type models
  • Thinking model — built on Qwen3-4B-Thinking for chain-of-thought reasoning
  • Lightweight deployment — 4B scale for consumer-hardware inference

Why It Might Matter

RL-trained offensive sub-agents represent a shift from static SFT toward environment-reward optimization:

  1. RL for vuln hunting — models trained against simulated pentest environments rather than static instruction corpora
  2. Eval infrastructure — Prime Intellect strix-xss environment enables reproducible XSS agent scoring
  3. PoC-to-production path — Kanti Labs documents SFT baselines (8B) and RL experiments (4B) as an iterative research line

For defenders and researchers, the value is understanding whether RL closes the gap between specialized XSS agents and human triage on structured eval harnesses.

Research Questions

  • How much does RL improve over the SFT Strix XSS line on identical eval environments?
  • Does the 4B RL model generalize beyond Prime Intellect simulated targets?
  • What Strix framework modifications are required for reliable sub-agent orchestration?

Supply Chain Note

Download from Hugging Face only. Verify checksums, prefer safetensors, and never load pickle-serialized weights in production pipelines. Third-party security models may carry elevated supply-chain risk — treat as untrusted code until audited.

Sources

Attribution

Open-source model indexed for security research inquiry. Uploaded to Hugging Face by kusonooyasumi under the Kanti Labs research program. OFFSITE.DARK does not endorse or distribute this artifact.