privesc-llm-rl-4b-itthipon

privesc-llm-rl:4b (GGUF, Q4_K_M) — Linux Privilege Escalation Specialist

First public full-model release. The upstream authors publish LoRA adapters only (sailab-vienna/privesc-llm-4b has no GGUF and no merged checkpoint), so this is the first public, serving-ready (merged + quantized) version of PrivEsc-LLM 4B (RL). Verification: HF model search for privesc-llm with the gguf filter returns this repo only. (Search-based evidence, not an exhaustive audit — but no other full release is currently publishable on HF.)

This is the Q4_K_M quantization of the merged PrivEsc-LLM 4B (RL) model, built for locally-run, authorized, automated Linux privilege-escalation assessment. It is the same weights evaluated in the paper below, quantized to fit consumer GPUs (≈2.6 GB + 8k-context KV cache fits the 4 GB class).

Author credit

Philipp Normann, Andreas Happe, Jürgen Cito, and Daniel Arp — "Towards Reliable Local Security Agents: Verifiable Post-Training for Linux Privilege Escalation" (arXiv:2603.17673, NDSS 2026), Security & AI Lab (SAILAB), TU Wien. Please cite this paper when using the model in research or published reports.

Base: Qwen/Qwen3-4B-Instruct-2507 (Apache-2.0) LoRA adapter: sailab-vienna/privesc-llm-4b (MIT)

What this model is (and isn't)

• It is: the paper's RLVR-trained 4B specialist for Linux privilege escalation — 93.3% success on the paper's 12-scenario held-out benchmark within 20 rounds.
• It is not: a general-purpose chatbot, nor an authorization to run any assessment without the explicit, written permission of the target system's owner.

Provenance / how this GGUF was built

Usage (Ollama)

ollama create privesc-llm-rl:4b -f - <<'EOF'
FROM privesc-llm-4b-rl-Q4_K_M.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20
PARAMETER num_ctx 8192
PARAMETER num_predict 2048
EOF

The specialist protocol (<tool_call>/<tool_response>, exec_command + test_credentials schemas) is injected by the client — the paper's privesc-protocol implementation from LONLY's models/privesc_protocol.py is byte-aligned with the training distribution.

Legal & safety

Authorized assessment only. This model generates privilege-escalation hypotheses for systems you own or are contracted to test. Unauthorized access to computer systems is illegal (e.g., the Computer Fraud and Abuse Act in the US, §33 of the Thai Cybercrime Act), so the upstream paper authors and re-distributors accept no liability for misuse.

Note on quantization: ollama/llama.cpp Q4_K_M retains ≈98%+ of base perplexity, so expect ≈90–95%+ of the paper's benchmark success vs its bf16 evaluation (paper runs the same weights unquantized on server GPUs).

Link / license

• Weights in this repo: Q4_K_M quantization of the merged model → Apache-2.0
• Adapter (upstream): MIT
• Base model: Apache-2.0