by Zach Stein-Perlman·Latest: September 2025 (frozen)·Source ↗
Weighted scorecard across seven categories (risk assessment, scheming prevention, safety research, misuse prevention, security, info sharing, planning).
Rows are organizations; columns are the dimensions AI Lab Watch scores. Cell text is the published grade (overall first, then per-pillar).
| Organization | Overall | Boosting safety research | Misuse prevention | Planning | Prep for extreme security | Risk assessment | Risk info sharing | Scheming risk prevention |
|---|---|---|---|---|---|---|---|---|
| Anthropic | Weak | Moderate | Very Weak | Very Weak | Very Weak | Weak | Weak | Very Weak |
| DeepSeek | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak |
| Google DeepMind | Very Weak | Moderate | Very Weak | Weak | Very Weak | Weak | Very Weak | Very Weak |
| Meta AI (FAIR) | Very Weak | Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak |
| Microsoft AI | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak |
| OpenAI | Very Weak | Weak | Very Weak | Very Weak | Very Weak | Weak | Weak | Very Weak |
| xAI | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Very Weak | Weak | Very Weak |
| Wave | Published | Orgs | License | Links |
|---|---|---|---|---|
| September 2025 (frozen)latest | 2025-09-01 | 7 | fair-use-citation |
Weighted scorecard across seven categories (risk assessment, scheming prevention, safety research, misuse prevention, security, info sharing, planning).
AI Lab Watch (ailabwatch.org) — frozen as of September 2025 per the author. Seven dimensions × seven frontier labs; per-cell scores in 0-100 percent. Overall is the site's published weighted total.
Grades are mirrored from upstream sources. We do not score organizations ourselves — see the source link for full methodology and per-indicator detail. Citation of published grades is consistent with fair-use attribution.