Written and maintained by CASRAI Editorial Board
Last updated
SaferAI is not the only group grading frontier AI companies on safety. The Future of Life Institute (FLI) — the nonprofit best known for the 2023 open letter calling for a pause on giant AI experiments — publishes its own AI Safety Index, and it asks a different question than SaferAI’s Tracker does.
This guide covers who publishes the Index, what it actually rates, and how it differs from the SaferAI rubric already covered in this cluster.
Who publishes it
The AI Safety Index is published by the Future of Life Institute. Unlike SaferAI’s Tracker, which is scored by SaferAI’s own methodology team, FLI convenes an independent panel of outside AI researchers and governance experts to review the evidence and assign the grades. The Summer 2026 edition’s panel included seven reviewers: David Krueger (University of Montreal), Sharon Li (University of Wisconsin–Madison), Tegan Maharaj (HEC Montréal), Sneha Revanur (Encode), Stuart Russell (UC Berkeley), Robert Trager (University of Oxford), and Yi Zeng (Renmin University of China).
FLI frames the project simply: AI experts rate leading AI companies on key safety and security domains. It has published editions in November 2024, July 2025 (Summer 2025), December 2025 (Winter 2025), and July 2026 (Summer 2026) — roughly twice a year since mid-2025.
What it rates
The Index grades each company across six domains. In the Summer 2026 edition, those domains and their indicator counts were:
- Risk Assessment (6 indicators)
- Current Harms (9 indicators)
- Safety Frameworks (4 indicators)
- Existential Safety (4 indicators)
- Governance & Accountability (4 indicators)
- Information Sharing (10 indicators)
That’s 37 indicators in total. The panel reviews company-specific evidence against each indicator and assigns a domain-level letter grade (A–F) based on absolute performance standards — a company is graded against what FLI considers adequate practice, not ranked against the other companies in the batch. Domain grades roll up into a single overall letter grade with a numeric score.
Two of the six domains are worth flagging because SaferAI’s rubric has no direct equivalent: Current Harms looks at problems a company’s deployed systems are already causing (not just documented risk process), and Existential Safety looks specifically at how a company’s plans address risks from future, more capable systems.
The Summer 2026 scores
Nine companies were rated in the July 2026 edition:
- Anthropic — C+ (2.66)
- OpenAI — C (2.28)
- Google DeepMind — C (2.01)
- Meta — D+ (1.32)
- Z.ai — D− (0.88)
- Alibaba Cloud — D− (0.87)
- xAI — F (0.65)
- DeepSeek — F (0.47)
- Mistral — F (0.33)
As with SaferAI’s Tracker, these numbers move between editions and FLI’s own site is the source of record for current figures — treat any fixed list, including this one, as a snapshot of the Summer 2026 edition rather than a live score.
How it differs from the SaferAI rubric
Both efforts independently grade frontier AI companies on safety-related practice, and it’s easy to conflate them because they cover overlapping companies. They are separate organizations with separate methodologies, and their scores are not interchangeable:
- Publisher and graders. SaferAI is scored in-house by SaferAI’s own research team. FLI’s Index is scored by an outside panel of academic and governance experts convened for that purpose.
- Scope. SaferAI’s four dimensions (risk identification, risk analysis and evaluation, risk treatment, risk governance) all measure the maturity of a company’s documented risk-management process. FLI’s six domains include that kind of process review (Safety Frameworks, Governance & Accountability) but also grade things SaferAI doesn’t directly score, such as harms already occurring in deployed systems (Current Harms) and how much a company discloses about its work (Information Sharing).
- Scoring format. SaferAI publishes a continuous 0–100% score per dimension. FLI assigns a letter grade (A–F) per domain, based on absolute standards rather than a curve, which then rolls up to an overall letter grade.
- What a low score means. A low SaferAI score flags specific gaps in what a company has published about its risk-management process. A low FLI grade can reflect that, or it can reflect harms or disclosure gaps that have nothing to do with how well-documented the company’s process is.
Neither is a substitute for the other. Consulting both gives a broader picture of a company’s safety practice than relying on either alone.
How to read it
- Check the edition date. FLI revises its indicators between editions — the domain names have stayed stable, but indicator counts and weighting have changed release to release, so grades from different editions aren’t always a clean like-for-like comparison.
- A grade is not a safety guarantee. As with SaferAI’s Tracker, the Index measures a company’s practice and disclosure against a rubric, not the real-world safety of any specific deployed system.
- Treat it as one input among several. The Index, SaferAI’s Tracker, and this cluster’s other framework guides each measure something different. None of them alone tells you whether a company’s AI systems are safe to use.
Frequently asked questions
Who publishes the AI Safety Index?
The Future of Life Institute. Grades are assigned by an independent panel of outside AI researchers and governance experts that FLI convenes for each edition, not by FLI staff directly.
Is the AI Safety Index the same as the SaferAI Tracker?
No. They are run by different organizations, using different methodologies, domains, and scoring formats. Both independently grade frontier AI companies on safety practice, but the scores are not directly comparable.
How many companies does the Index grade?
Nine, as of the Summer 2026 edition (July 2026): Anthropic, OpenAI, Google DeepMind, Meta, Z.ai, Alibaba Cloud, xAI, DeepSeek, and Mistral. Earlier editions rated fewer companies.
How often is it updated?
FLI has published four editions so far: 2024, Summer 2025, Winter 2025, and Summer 2026 — roughly twice a year since mid-2025.
Where can I see the current, exact scores?
FLI publishes the full report, methodology, and per-company rationale on its own site for each edition. That is the authoritative source for current figures rather than any fixed summary.
Related on CASRAI
This guide is part of CASRAI’s frontier AI safety and governance cluster. For the SaferAI rubric this Index is most often compared against, see How to Grade a Frontier AI Safety Framework: The SaferAI Rubric. For the elements that structure how organizations document AI system provenance and safety claims, see NIKOLAI.
