qwen3.8-27b-cyber-exploit-agent

## https://sploitus.com/exploit?id=7B22971A-BFDC-5C14-AEB1-5123B89358A8
---
library_name: peft
base_model: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
license: apache-2.0
language:
  - en
tags:
  - security
  - code
  - vulnerability-analysis
  - xrpl
  - rippled
  - c-plus-plus
  - exploit-development
  - qlora
  - lora
  - gguf
  - uncensored
  - grpo
---

# Qwen3.8-27B Cyber Exploit Agent (AEON base)

Offensive/defensive vulnerability-analysis fine-tune of
[AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16](https://huggingface.co/AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16).
The abliteration ships in the base — this fine-tune does **not** re-abliterate and does **not**
train refusal directions back (per training contract). Phase-0 refusal smoke on the raw base:
0/10 hard refusals on offensive analysis prompts.

## Pipeline (per contract)

1. **Phase 0 — diagnostics:** native chat template verified (assistant turns split into
   `reasoning_content` + content, generation prompt ends in `\n`); refusal smoke 0/10;
   QLoRA VRAM pretest @3072 peak 27.5 GiB on a 24 GB RTX 5090 laptop (WSL2 oversubscription —
   documented, not hidden) → max_seq stays 3072, no 4096 upgrade.
2. **Phase 1 — SFT:** QLoRA NF4 double-quant, LoRA r16/α16 on q/k/v/o + gate/up/down (no lm_head),
   79.7 M trainable (0.30 %), batch 1 × grad-accum 16, lr 2e-5 cosine, 2 epochs max with
   best-checkpoint-by-eval-loss selection, assistant-only loss mask (common-prefix diff).
   Dataset v4: 451 code-verified samples — 280 CyberGym train issues (blacklist-cleaned, no eval
   leakage), XRPL findings pinned to real `rippled` code, 18 self-built memory-safety labs with
   ASan/UBSan evidence (shipped in `training_data/`).
3. **Phase 2 — RL (laptop-honest scale):** dfs-large1-style reward math on 16 episodes
   (declared, not inflated): R_verif +2.0 sandbox-executable trigger / +1.0 verdict-gold / +0.5
   parseable TRIGGER / +0.5 honest writeup; R_effort log-penalties on tokens/idle; R_report
   penalty for unverified findings; R_refuse −3.0; R_wrong −1.5; R_crash −0.2; clipped [−5,+5].
   REINFORCE with EMA baseline (group=1 on 1×5090), KL anchor β=0.02 vs the frozen SFT adapter,
   lr 5e-7. Sandbox: `unshare -n` (no network) + uid drop to nobody + rlimits. 16 real episodes,
   not a fake 100 — this is 1×5090 reality, not Fireworks scale.
4. **Phase 3 — gates:** eval vs the raw AEON base; RL discarded if worse than SFT (per contract).

## Results

(see `eval_27b/` logs shipped below — honest, including failures)

EVAL_PLACEHOLDER

## Output contract

`` white-box reasoning → `### TRIGGER` (minimal reproducer) → `### EXPLOIT WRITEUP` →
`### VERDICT` with one of `VALID_BUG [TRACK:ledger|TRACK:safety PATTERN:N11-N18]`,
`FALSE_POSITIVE`, `HYGIENE`, `CORRECTNESS-ISSUE`, `UNPROVEN`, `INSUFFICIENT_EVIDENCE`.

## Usage (LM Studio / llama.cpp)

`--jinja`, thinking on, temperature 0.6 / top_p 0.95 / top_k 20 (XRPL triage: 0.2–0.4 +
majority-of-3 + human review), max_tokens ≥ 1200. System prompt: `inference_system.txt`.
The GGUF includes the MTP (speculative decoding) tensors required by current llama.cpp qwen35
support.

## Files

- `Qwen3.8-27B-Cyber-Exploit-Q4_K_M.gguf` (recommended) / `Q5_K_M.gguf`
- `inference_system.txt` — the system prompt used at training and eval
- `training_data/` — the exact dataset (`train_all_v4.jsonl`), labs, ASan/UBSan evidence,
  dataset gate report, all build/train/RL/eval scripts

## Honest limitations

- RL scale is 16 episodes on one laptop GPU. The reward math follows the dfs-large1 contract,
  but the training volume is not comparable to datacenter RL runs. Judge the SFT layer as the
  main artifact.
- XRPL verdicts remain sampling-sensitive (temp 0.6 vs 0.2 differ; per-draw variance). Use low
  temperature + majority voting + human review.
- No Mythos-869 or frontier-model superiority claims — comparisons here are vs the same base
  model before fine-tuning.
- WSL2 GPU memory oversubscription (>24 GB allocations spill to system RAM): training completed,
  but step times are slower than native VRAM fitting.

Intended for defensive security research, CTF, and analysis of systems you own or are authorized
to test.