Overview
Developers
What each developer publishes about the safety of its own models, side by side: its documents, the values recorded from them, whether they keep a changelog, and which metric families they report.
Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog
What each developer reports
The dataset holds 70 documents from 8 developers, covering 89 models, with 1,645 values in all. Developers publish different kinds of documents and run different tests, so the counts on this page describe what each developer reports, not how safe its models are.
The chart shows, for each developer, the share of its documents with at least one printed number in each metric family.
Metric families reported, by developer
Share of each developer's documents with at least one printed number in each metric family. Developers are ordered by number of documents (in brackets), then by name. A cell opens the data explorer at that family and developer.
| Developer | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 | M12 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OpenAI (22) | 18% | 5% | 18% | 45% | 41% | 5% | 86% | 45% | 55% | 59% | 0% | 18% |
| Google DeepMind (16) | 38% | 0% | 31% | 0% | 13% | 0% | 88% | 13% | 6% | 38% | 0% | 6% |
| Anthropic (15) | 60% | 53% | 53% | 53% | 60% | 13% | 80% | 13% | 67% | 87% | 7% | 20% |
| xAI (8) | 13% | 0% | 13% | 0% | 100% | 88% | 100% | 88% | 63% | 100% | 0% | 0% |
| Meta (5) | 40% | 20% | 40% | 40% | 40% | 40% | 40% | 40% | 60% | 60% | 0% | 0% |
| Moonshot AI (2) | 0% | 0% | 0% | 0% | 0% | 0% | 50% | 50% | 50% | 50% | 0% | 0% |
| DeepSeek (1) | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% |
| Zhipu AI (1) | 0% | 0% | 0% | 0% | 0% | 0% | 100% | 0% | 0% | 0% | 0% | 0% |
Share of documents0%1-24%25-49%50-74%75-100%
- M1
- Evaluation awareness
- M2
- Reward hacking
- M3
- Sabotage and sandbagging
- M4
- Misalignment audits
- M5
- Honesty and hallucination
- M6
- Sycophancy
- M7
- Harmful compliance and over-refusal
- M8
- Jailbreak robustness
- M9
- Prompt injection
- M10
- Dangerous capabilities and risk determinations
- M11
- Self-preservation
- M12
- Chain-of-thought monitorability
The chart is already a table: every cell prints its share, and its label gives the two counts behind it. The CSV has the counts.
Read this as a lower bound: some documents were not read to the end in version 0 (each document's page says how far), so families reported late in them may be undercounted.
A document counts for a family when it prints at least one number in it. Values read off a figure, stated in words, or given as a category such as a risk level do not count.
Source: Safety Card Ledger v0.1 · Data from developer system cards · CC BY 4.0
Side by side
In name order. Families reported: the metric families in which at least one of the developer's documents prints a number. Counts describe what each developer publishes, not the safety of its models.
| Developer | Documents | Models | Values | Families reported |
|---|---|---|---|---|
| Anthropic15 documents · 20 models · 512 values · 12 of 12 families reported14 system cards and 1 system card addendum22 May 2025 to 22 Sep 2026 | 15 | 20 | 512 | 12 of 12 |
| DeepSeek1 document · 1 model · 1 value · 0 of 12 families reported1 technical report17 Sep 2025 | 1 | 1 | 1 | 0 of 12 |
| Google DeepMind16 documents · 14 models · 175 values · 8 of 12 families reported12 model cards, 2 Frontier Safety Framework reports, 1 technical report and 1 launch post17 Jun 2025 to 2 Sep 2026 | 16 | 14 | 175 | 8 of 12 |
| Meta5 documents · 5 models · 37 values · 10 of 12 families reported2 model cards, 2 safety reports and 1 launch post5 Apr 2025 to 10 Aug 2026 | 5 | 5 | 37 | 10 of 12 |
| Moonshot AI2 documents · 2 models · 5 values · 4 of 12 families reported2 technical reports28 Jul 2025 to 27 Jul 2026 | 2 | 2 | 5 | 4 of 12 |
| OpenAI22 documents · 38 models · 800 values · 11 of 12 families reported16 system cards, 3 system card addenda, 1 model card, 1 launch post and 1 policy post16 Apr 2025 to 16 Sep 2026 | 22 | 38 | 800 | 11 of 12 |
| xAI8 documents · 8 models · 113 values · 8 of 12 families reported1 system card and 7 model cards20 Aug 2025 to 21 Sep 2026 | 8 | 8 | 113 | 8 of 12 |
| Zhipu AI1 document · 1 model · 2 values · 1 of 12 families reported1 technical report8 Aug 2025 | 1 | 1 | 2 | 1 of 12 |
Changelog practice
A document has a changelog when it keeps its own record of the changes made to it after publication, such as a dated list of corrections. Where a document keeps none, we find changes by comparing its versions (Methodology §7).
Of the 70 documents, 15 have a changelog, 4 have a partial changelog, 20 have no changelog, for 17 it is not known whether they keep one and for 14 a changelog does not apply.
- Has a changelog
- The document keeps its own record of the changes made to it after publication.
- Partial changelog
- The document records some of its changes, but not all of them, or not in full.
- No changelog
- The document keeps no record of its changes.
- Not known
- We have not yet established whether the document keeps a record of its changes.
- Not applicable
- Our source registry records a changelog as not applying to the document, for example a research paper or a launch post.
| Developer | Documents | Has a changelog | Partial changelog | No changelog | Not known | Not applicable |
|---|---|---|---|---|---|---|
| Anthropic15 documents · Has a changelog: 11 · No changelog: 1 · Not known: 3 | 15 | 11 | 0 | 1 | 3 | 0 |
| DeepSeek1 document · Not applicable: 1 | 1 | 0 | 0 | 0 | 0 | 1 |
| Google DeepMind16 documents · No changelog: 12 · Not applicable: 4 | 16 | 0 | 0 | 12 | 0 | 4 |
| Meta5 documents · Partial changelog: 1 · Not applicable: 4 | 5 | 0 | 1 | 0 | 0 | 4 |
| Moonshot AI2 documents · Not applicable: 2 | 2 | 0 | 0 | 0 | 0 | 2 |
| OpenAI22 documents · Has a changelog: 4 · Partial changelog: 2 · Not known: 14 · Not applicable: 2 | 22 | 4 | 2 | 0 | 14 | 2 |
| xAI8 documents · Partial changelog: 1 · No changelog: 7 | 8 | 0 | 1 | 7 | 0 | 0 |
| Zhipu AI1 document · Not applicable: 1 | 1 | 0 | 0 | 0 | 0 | 1 |