FLI AI Safety Index
by Future of Life InstituteLetter-grade ratings of frontier AI labs across six safety domains: risk assessment, current harms, safety frameworks, existential safety, governance, and information sharing.
External Scorecards
Side-by-side overall grades from the five major external AI-safety scorecards. Click any row to view per-dimension grades on the organization's profile.
Scorecards
5
Sources w/ Data
5
Organizations
25
Cells
57
Overall grades — latest wave per scorecard
| Organization | FLI Index | SaferAI | AI Lab Watch | FMTI | Seoul Tracker |
|---|---|---|---|---|---|
| AI21 Labs | — | — | — | 66 | — |
| Alibaba Cloud | D- | — | — | 26 | — |
| Amazon | — | Very Weak | — | 39 | Fulfilled |
| Anthropic | C+ | Weak | Weak | 46 | Fulfilled |
| Cohere | — | Very Weak | — | — | Partial |
| DeepSeek | D | — | Very Weak | 32 | — |
| G42 | — | Very Weak | — | — | Fulfilled |
| — | — | — | 41 | Fulfilled | |
| Google DeepMind | C | Very Weak | Very Weak | — | — |
| IBM | — | — | — | 95 | Unfulfilled |
| Inflection AI | — | — | — | — | Unfulfilled |
| Magic | — | Very Weak | — | — | — |
| Meta AI (FAIR) | D | Very Weak | Very Weak | 31 | Partial |
| Microsoft AI | — | Very Weak | Very Weak | — | Fulfilled |
| Midjourney | — | — | — | 14 | — |
| Mistral | — | — | — | 18 | Unfulfilled |
| Naver | — | Very Weak | — | — | Partial |
| NVIDIA | — | Very Weak | — | — | — |
| OpenAI | C+ | Weak | Very Weak | 35 | Fulfilled |
| Samsung Electronics | — | — | — | — | Unfulfilled |
| Technology Innovation Institute | — | — | — | — | Unfulfilled |
| Writer | — | — | — | 72 | — |
| xAI | D | Very Weak | Very Weak | 14 | Partial |
| Z.ai | D | — | — | — | — |
| Z.ai | D | — | — | — | Unfulfilled |
Methodology
SaferAI Ratings
by SaferAIContinuously-updated risk-management ratings for frontier developers across four pillars: risk identification, analysis & evaluation, treatment, and governance.
Weighted scorecard across seven categories (risk assessment, scheming prevention, safety research, misuse prevention, security, info sharing, planning).
Foundation Model Transparency Index
by Stanford CRFMTransparency-focused index scoring developers on 100 indicators across upstream resources, the model itself, and downstream use.
Seoul Commitment Tracker
by The Midas ProjectTracks adherence to the Seoul Frontier AI Safety Commitments via 'Fulfilled / Partial / Unfulfilled' verdicts on five red-line components.
Grades are mirrored from upstream sources. We do not score organizations ourselves — see each row's source link for methodology and full per-dimension grades. Citation of published grades is consistent with fair-use attribution.