How 382 GB of Proprietary Data Retaught Kimi K3.1 Pro Offensive Cyber
The 8/10 CyberGym result came from our open unrestricted Kimi K3.1 Pro model. Five training iterations on 382 GB of proprietary offensive-security data then produced a separate 9/10 run, matching OMEN on the same subset. Here is the model lineage, what changed, and which weights are open.
For authorized security research only. CyberGym evaluates reproductions of known, already-fixed vulnerabilities in instrumented targets. It does not authorize testing against live third-party systems.
This post does two things. First, it lays out the model lineage behind our earlier CyberGym benchmark report. Second, it explains what happened after we continued training that checkpoint on 382 GB of proprietary offensive-security data.
The short version: the model that scored 8/10 was our open unrestricted Kimi K3.1 Pro checkpoint. Its performance was already comparable to our other unrestricted GLM-5.3 model. Five iterations of domain-specific training then produced a specialized checkpoint that scored 9/10 on the same ten-task CyberGym subset, matching OMEN's earlier score.
Those are different checkpoints and different claims. Keeping them separate matters.
The CyberGym model lineage
The CyberGym model lineage became explicit as we continued developing the family. WARLOCK now points to a specialized closed-weight checkpoint. The 8/10 run belongs to penclaw-Kimi-K3.1-pro-unrestricted, our open-weight Kimi K3.1 Pro derivative. The resulting ordering is:
| CyberGym result | Model attribution |
|---|---|
| 9/10 | Fine-tuned penclaw-Kimi-K3.1-pro-unrestricted — specialized closed weights; separate post-training run |
| 9/10 | OMEN — separately benchmarked checkpoint |
| 8/10 | penclaw-Kimi-K3.1-pro-unrestricted — open weights; original pre-training run |
| 7/10 | apex-flash-1-abliterated |
| 6/10 | Venice abliteration-abliterated-model-large-v2 |
The original per-task outcomes attributed to the 8/10 run remain attached to that unrestricted Kimi checkpoint: it failed arvo:368 and oss-fuzz:42535468. The model-lineage mapping does not retroactively rewrite the historical task matrix.
The post-training run is a separate result. Its 9/10 score means the fine-tuned checkpoint solved arvo:368, the FreeType CFF2 blend task the open checkpoint had missed, while oss-fuzz:42535468 remained the sole miss. That is the measurable one-task improvement: the OpenSC pkcs15init key-length bug continued to defeat every model, but the fine-tuned Kimi checkpoint joined OMEN in clearing the FreeType heap-use-after-free task.
The benchmark's execution and verification protocol remains as previously described: a submission counted only when the target actually crashed, and the ground-truth evidence was the PoC server's exit-code log. Model attribution is metadata; it does not change what the target did.
The baseline was already strong
One popular assumption around frontier-model releases is that weak offensive-cyber behavior must mean the useful cyber data was removed during training. That explanation was too tidy for what we observed.
Our unrestricted Kimi K3.1 Pro checkpoint scored 8/10 on the ten-task CyberGym subset. That is not the profile of a model whose entire security-research capability had been emptied out. It was performing in the same broad territory as our other unrestricted GLM-5.3 model.
This observation does not reconstruct the original training corpus, and we cannot prove a negative about every dataset that may or may not have been included. What the run did establish was narrower and more useful: the open unrestricted checkpoint still had substantial offensive-cyber capability to build on.
That distinction shaped the next step. The problem was no longer simply whether the model would engage with security material. It was whether targeted training could improve how it reasoned through an engagement, pursued a promising signal, and closed the loop with a working proof.
Five iterations on 382 GB of proprietary data
We used 382 GB of proprietary offensive-security data for continued training and ran five training iterations.
This was not a cosmetic system-prompt change. It was domain-specific fine-tuning intended to strengthen the behaviors that matter in autonomous security work: exploit reasoning, debugging failed attempts, pivoting after negative results, and producing proof rather than merely describing a possible vulnerability.
The resulting checkpoint scored 9/10 on the same CyberGym subset, tying OMEN's earlier result on that benchmark. The one-task gain came from solving arvo:368; the specialized checkpoint did not clear the universal OpenSC wall. Separately, models in this performance class have previously measured up to 98% pass@1 on our applicable offensive-security evaluations. That figure is not a percentage derived from the ten-task CyberGym run, and the two evaluation contexts should not be arithmetically combined.
The practical result is the important part: targeted post-training produced a materially stronger offensive-cyber model from a foundation that was already capable. On the tasks where the two Kimi checkpoints differ, the specialized model succeeded on the FreeType CFF2 blend heap-use-after-free where its open precursor had failed.
The gain was therefore not "restore a model from zero." It was specialize an already-strong checkpoint until it behaved like our strongest offensive models.
Abliteration was not the whole story
The explicit lineage changes the interpretation of the earlier benchmark.
An open checkpoint with fewer operational guardrails had already reached 8/10. Removing obstacles to engagement did not, by itself, explain its ceiling. The subsequent improvement came from deliberate domain training on proprietary offensive-security material.
That is consistent with a broader lesson from the original runs: weights and training data determine more than a prompt or a refusal policy. Guardrails can affect whether a model attempts the work, but task-specific capability still depends on what the model has learned to do.
For us, the practical equation was:
Open unrestricted foundation
+ 382 GB of proprietary offensive-security data
+ five training iterations
= a specialized checkpoint scoring 9/10 on the same CyberGym subset
That is also why we describe the result as teaching hacking intuition, rather than merely removing a refusal behavior. The relevant shift was in execution: investigating, adapting, and proving.
The three versions, plainly
Model names are only useful if their lineage is explicit. Here is the release status as it stands:
| Version | Lineage and purpose | Weights |
|---|---|---|
| penclaw-Kimi-K3.1-pro-unrestricted | Open unrestricted Kimi K3.1 Pro derivative; the source of the earlier 8/10 result | Open on Hugging Face |
| Necromicon | Proprietary offensive-cyber checkpoint derived through further domain training | Closed |
| Warlock | Proprietary offensive-cyber checkpoint in the same release family | Closed |
Necromicon and Warlock are the specialized versions. They are not new names for the open Hugging Face checkpoint, and the open checkpoint is not being presented as the repository for the later proprietary fine-tunes.
If you are new to the project, the open foundation is available now at huggingface.co/audnai/penclaw-Kimi-K3.1-pro-unrestricted. The specialized Necromicon and Warlock weights remain closed. That is a deliberate release boundary, not a claim that the open model contains the subsequent 382 GB training run.
Why the specialized weights are staying closed
When a specialized checkpoint performs in the class we associate with GPT-6-astra on the relevant offensive workloads, the training data and resulting weights become substantial research assets.
The open unrestricted model still lets researchers inspect and build on the foundation. Necromicon and Warlock remain closed while we continue evaluating and deploying their specialized capabilities. We have not yet published those weights on Hugging Face, and this post does not announce a future release date.
We think that distinction is more useful than a vague launch promise: the precursor is open; the two specialized checkpoints are closed; the 382 GB corpus is proprietary.
What this update does—and does not—claim
A precise result is easier to test than a broad one.
What we are saying:
- The earlier 8/10 CyberGym result came from the open unrestricted Kimi K3.1 Pro checkpoint.
- Its baseline performance was comparable to our other unrestricted GLM-5.3 model.
- We subsequently used 382 GB of proprietary offensive-security data over five training iterations.
- The specialized checkpoint subsequently scored 9/10 on the same CyberGym subset, matching OMEN's previous tally.
- Its remaining miss was
oss-fuzz:42535468; the separate post-training run solvedarvo:368. - Necromicon and Warlock remain closed-weight releases, while the precursor remains available on Hugging Face.
What we are not saying:
- That the old 8/10 run should be reclassified as a new post-training run.
- That failures in the original per-task matrix have silently become passes.
- That the 98% pass@1 figure is a percentage derived from the ten-task CyberGym subset.
- That the open Hugging Face checkpoint includes the later proprietary offensive-cyber fine-tune.
- That unrestricted behavior alone was sufficient to produce the final specialized capability.
Conclusion
The important story is not that a model went from incapable to capable overnight. The baseline was already strong: 8/10 on a demanding, crash-verified vulnerability-discovery subset.
The new result is that five iterations on 382 GB of proprietary offensive-security data made that foundation materially stronger: the specialized checkpoint reached 9/10, matching OMEN's earlier tally on the same subset. It solved the FreeType task its predecessor missed and still failed the OpenSC task every model missed. Keeping the model lineage explicit makes that progression visible: open unrestricted precursor first, specialized Necromicon and Warlock checkpoints second.
The open precursor remains on Hugging Face. The specialized weights remain closed. Both facts can be true without blurring which model produced which result.
Questions about the model lineage or release status can be sent to support@audn.ai.