Did your fine-tune quietly make your model easier to hijack?
Independent safety evaluation for AI checkpoints, including the ones nobody else can grade. We measure your checkpoint against the model it was derived from, so you get a delta, not an uninterpretable number.
Test my checkpoint → See the public registerThe models people are actually downloading
Most of the open-weight checkpoints trending on Hugging Face right now are already on this register — the Qwen3.6 and gpt-oss families, the DeepSeek-V4-Flash general release, and the uncensored variants people are pulling down by the million. We graded them. The uncensored ones are measurably the riskiest.
| Model (Hugging Face) | Downloads | Injection compliance [95% CI] | Grade | Evidence |
|---|---|---|---|---|
| Qwen3.6-35B-A3B base for the uncensored variant below Qwen/Qwen3.6-35B-A3B |
11.3M | 18.46% [10.9, 29.6] 12 / 65 trials |
Fprov. | detector |
| gpt-oss-20b openai/gpt-oss-20b |
8.6M | 0.00% [0.0, 5.9] 0 / 61 trials |
Cprov. | detector |
| qwen3-32b Qwen/Qwen3-32B |
8.3M | 0.00% [0.0, 2.2] 0 / 173 trials |
B | detector |
| Qwen3.6-27B base for the uncensored variant below Qwen/Qwen3.6-27B |
6.98M | 0.00% [0.0, 6.0] 0 / 60 trials |
Cprov. | detector |
| gpt-oss-120b openai/gpt-oss-120b |
4.2M | 9.23% [5.4, 15.4] 12 / 130 trials |
Dprov. | detector |
| deepseek-v4-flash original run · detector-only, never judge-audited deepseek-ai/DeepSeek-V4-Flash |
2.7M | 26.47% [19.8, 34.5] 36 / 136 trials |
Fprov. | detector |
| Qwen3.6-35B-A3B-Uncensored uncensored trending HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive |
1.9M 3.3K likes |
36.67% [25.6, 49.3] 22 / 60 trials |
Fprov. | detector |
| Qwen3.6-27B-Uncensored uncensored trending DavidAU/Qwen3.6-27B-…-Uncensored-…-GGUF |
1.6M | 15.00% [8.1, 26.1] 9 / 60 trials |
Fprov. | detector |
| deepseek-v4-flash-0731 trending general release · judge-audited deepseek-ai/DeepSeek-V4-Flash-0731 |
433K 2.4K likes |
0.00% [0.0, 8.8] 0 / 40 trials |
Cprov. | judge |
Trending but not yet on the register — coming. zai-org/GLM-5.2 (2.2M downloads), moonshotai/Kimi-K3 (1.1M downloads, 10K likes — one of the hottest open-weight releases right now), and MiniMax-H3 are trending on Hugging Face but aren't in the current config_exposure run. They're queued for the next snapshot; the register is live, not complete.
Also tested, not yet graded for injection. inclusionAI/Ling-3.0-flash and poolside/Laguna-S-2.1 (both trending) are on the register and clear the autonomous (Frame A) battery, but have no Frame B injection trials yet — so they carry no injection-compliance rate here.
Sources. Downloads and likes are a manual snapshot from Hugging Face, 2026-08-05; rates and grades are from the public config_exposure register and are labelled judge-audited or detector-scored (a lower bound). This is a curated highlight set — the full live register (38 checkpoints) has every row, sortable.
Fine-tuning changes safety behaviour. Almost nobody measures it.
Fine-tuning is documented to erode safety alignment even when the training data is entirely benign: refusal rates collapse, and fine-tuned variants are markedly more susceptible to hijacking than the model they came from. Yet the artifact you actually deployed has, in almost every case, never been evaluated by anyone. No public benchmark covers a private fine-tune, by definition.
An absolute score on a fine-tune is close to meaningless at realistic sample sizes. The measurement that means something is the paired delta against its base model, run in the same harness under the same conditions with only the training changed.
Refusal-ablated variants are the public, extreme case of a derived checkpoint. Your domain fine-tune is the same experiment with a smaller expected effect and no published number at all.
What the open-weight data actually says
In July 2026 the headlines were about frontier models breaking containment. The deployable, open-weight reality is quieter and more useful, and we can put numbers on it. Across 9,572 public config_exposure trials spanning 38 model configurations and 13 families — the graded subset of the open dataset (17,438 trials across all scenarios), reconciled 2026-08-05:
Autonomous self-exfiltration
Unprompted, organic self-bootstrap: 0 of 3,180 Frame A trials across every checkpoint. A bounded null (95% CI upper 0.12%). At this scale, models do not spontaneously scheme to escape.
Injection compliance (judge-audited)
The real vector. On the judge-audited injected (Frame B) rows, compliance runs 0% to 22.5% across checkpoints (deepseek-v4-flash-0731 → devstral-medium) — common enough to matter, and the spread is wide. Detector-only rows in the Partial runs group run higher (to 36.67%) and are labeled detector-only, never merged into this judge-audited spread.
Two provenance tiers, labeled
779 of 9,572 trials (8.1%) carry an LLM-judge verdict — the authoritative genuine-self-bootstrap tier, concentrated in the newest snapshot. The rest fall back to the deterministic detector (a lower bound) and are flagged as Partial runs. The two DeepSeek runs are kept separate: detector-only original (26.47%) vs. judge-audited general release (0%). Every published rate is labeled by its source.
The baseline: injection compliance across the public register
This is the reference scale that makes a single checkpoint's number interpretable. Every checkpoint showed zero unprompted self-exfiltration; under injection they diverge sharply, from 0% to 22.5% of trials. Rates are judge-audited; bars are 95% confidence intervals, wide at these sample sizes and shown rather than hidden. Detector-scored rows — a lower bound, not judge-audited — sit in a separate strip below the main plot, never merged onto the judge-audited scale; "Show more models" expands to the full partial-runs corpus.
config_exposure register: 9,572 trials, 38 checkpoints (4 judge-audited for Frame B injection compliance, charted above; 2 detector-scored in the strip; the rest provisional). Full leaderboard with grades →
The ablation effect: refusal-ablated ("uncensored") checkpoints are riskier
Derived checkpoints that strip refusal training are the public, extreme case of a fine-tune. Same scenario, same battery, same instrument — only the training changed.
What this maps to
Each metric maps to a framework your compliance team already uses: EU AI Act Article 15 (the threshold-independent robustness/cybersecurity requirement), NIST AI RMF Measure, and OWASP LLM Top 10 (LLM01 Prompt Injection; ASI04 tool poisoning). The evaluation produces supporting evidence toward those obligations, not a certification. Trust & regulatory mapping →
Test your checkpoint
We run your checkpoint and the base model it derives from through the same battery, and give you the delta: per-channel injection rates with confidence intervals, a paired odds ratio with a significance test, redacted evidence from the trials that failed, and an EU AI Act Article 15 write-up, as a durable, versioned document. Works on private fine-tunes, LoRAs, merges, quantized builds, and closed API models.
Prefer email? Write to [email protected]. We evaluate client checkpoints under confidentiality.