Unrestricted Kimi K3 Fine-tune 'CyberKimi' Aces ExploitBench V8 Bug Challenges
Anony6666 · reddit · 2026-08-10
Security researcher Taha (lordx64) fine-tuned Moonshot's Kimi K3 (2.8T MoE) with removed guardrails to create CyberKimi, a model designed specifically for red and blue team cybersecurity operations.
Tested on ExploitBench against a hardened Chrome V8 bug (CVE-2024-6100), the model showed strong potential:
- Stock Kimi K3 scored only 4/16.
- CyberKimi unassisted scored 8/16.
- With a methodology prompt pack, it reached 10/16.
This outperforms Claude Opus 4.7 and base GPT-5.5, sitting just behind the top-ranked Claude Mythos Preview (16). The author open-sourced the full chain-of-thought transcripts and grading data. The model is positioned for both exploit development (red team) and detection engineering (blue team).
More from Models
- Moonshot Defies Odds: Building Frontier Model Kimi K3 with ~500 Staff — OwainEvans_UK · 2026-08-10
- Mistral Launches Shieldstral: A 3B Parameter Open-Source Safety Classifier — dl_weekly · 2026-08-10
- Google SDK Leak Reveals 'Gemini 4 Flash' Reference — Rare_Bunch4348 · 2026-08-10
- OpenAI Models Hit 93% Hallucination Rates, Challenging AI Unit Economics — gerardsans · 2026-08-10
- Anthropic's Strategy: Focuses on B2B Coding Models, Skips Image Generation — sahilypatel · 2026-08-10
- 20VC Founder Says Qwen and Kimi Beat ChatGPT and Gemini for Research — hsu_byron · 2026-08-10