AI vendor safety has just been graded, and the report card is harsh. In its AI Safety Index published in July 2026, the Future of Life Institute assessed nine major artificial intelligence labs: none earned an A or a B, and the top grade stops at C+. For an SME leader handing documents, emails or code to one of these vendors, the point is not to panic, but to know what to do with this information.

Key takeaways

• The Future of Life Institute (FLI) published its AI Safety Index in July 2026, grading nine AI labs on their safety practices, assessed by a panel of 7 independent experts (including Stuart Russell and David Krueger).
• No company earns an A or a B. Anthropic leads with C+ (2.66/4), ahead of OpenAI (C, 2.28) and Google DeepMind (C, 2.01).
• Three companies fail with an F: xAI (0.65), DeepSeek (0.47) and Mistral (0.33). They come from the United States, China and Europe: the safety gap is global, not regional.
• The panel's central finding: safety pledges made during fundraising rounds have been quietly weakened under competitive pressure.
• For an SME, this index says nothing about model quality or GDPR compliance. It measures the maturity of a vendor's internal practices: one criterion among others, useful mainly to arbitrate between two equivalent options.

What the index actually measures

The AI Safety Index is an independent ranking published by the Future of Life Institute, a non-profit organisation focused on the risks of advanced technologies. It does not test the models: it assesses the declared and verifiable practices of the companies that build them, based on public documents and questionnaires sent to the labs.

The panel of 7 experts grades each company across six domains:

The full July 2026 ranking

Here are the overall grades published by the Future of Life Institute, on a 0 to 4 scale.

Anthropic leads five of the six domains, notably through its transparency and a more established safety framework. OpenAI takes the lead on the Risk Assessment domain alone, driven by a broader evaluation suite and more varied use of external testers (source: Future of Life Institute).

One point deserves European readers' attention: Mistral, the only French lab in the ranking, receives the lowest grade (F, 0.33). This does not mean its models are dangerous or non-compliant with GDPR. It means the company publishes little documentation on its safety practices, and the index penalises that silence heavily. A discreet lab scores poorly in an index that mostly measures what is made public.

The finding that matters: pledges walking back

The most notable element of the report is not the ranking itself, but a trend. The panel documents a consistent pattern: safety promises made during fundraising rounds were later softened or voided, under competitive pressure.

Two examples cited in the report:

• Anthropic, OpenAI, Google DeepMind and Meta all pulled back from their commitment to pause development if safety red lines were approached.
• Those same companies, which previously banned military applications of their models, gradually reversed that position.

The panel calls this a "movement of goalposts". On the Existential Safety domain, no company exceeds C- and current approaches, such as interpretability, are judged largely inadequate.

How an SME should read this ranking

The risk here is over-interpretation. This index measures the maturity of a vendor's internal practices against largely systemic risks. It says nothing about three things that matter far more day to day for an SME.

Put differently: a good grade does not remove the need for a solid contract, and a poor grade does not disqualify a vendor whose commercial and technical terms meet your need. The index is a governance signal, not a compliance audit.

It stays useful in three concrete situations:

1. Arbitrating between two vendors that are equivalent on price and performance. At comparable terms, practice maturity is a defensible tie-breaker.
2. Backing up a selection file. If you must justify a decision to a client, an insurer or an auditor, an independent and dated ranking is admissible evidence.
3. Asking the right questions. The index's six domains make an excellent questionnaire to send a vendor before signing.

The four checks that actually protect you

A safety ranking does not replace contractual points. Before entrusting company data to an AI tool, four checks remain the priority.

These four points appear in a contract and are enforceable. A grade in an index is not. This is why we recommend treating the AI Safety Index as a context indicator, alongside your own requirements, rather than as the deciding criterion.

The measured bright spot: this index exists, it is public, it is repeated and it is comparable over time. A sector willing to be graded publicly on its safety improves faster than an opaque one. Meta moved from 6th to 4th place between two editions, while xAI slipped from 4th to 7th: the ranking moves, so it carries weight.

FAQ

What is the Future of Life Institute's AI Safety Index?

It is an independent ranking that grades the safety practices of major AI labs. The edition published in July 2026 assesses nine companies across six domains, via a panel of 7 independent experts. It relies on the companies' public documents and on questionnaires sent to them.

Which company gets the best grade in 2026?

Anthropic, with C+ (2.66 out of 4). It is the highest grade in the ranking: no company earns an A or a B. OpenAI (C, 2.28) and Google DeepMind (C, 2.01) follow.

Should I avoid vendors graded F, such as Mistral or DeepSeek?

Not mechanically. The index measures transparency and the maturity of declared safety practices, not GDPR compliance or model quality. A low grade may reflect a lack of public documentation rather than a technical flaw. The decision should rest on your contractual requirements: training on your data, location, retention, reversibility.

Does this index have legal weight for my SME?

No. The Future of Life Institute is a non-profit organisation, not a regulator. This index creates no obligation and is not a certification. Your real obligations come from GDPR and the EU AI Act, whose deadlines keep approaching in 2026.

What to do with it

A ranking where the top grade is C+ deserves to be read for what it is: a thermometer for the maturity of a young sector, not a verdict on the tools. For an SME, the practical conclusion fits in one sentence: choose your vendor on your contract, and use the index to ask better questions before you sign it.

To go further on AI governance and tool selection, browse the other guides in the LUWAI Mag or see how our clients frame their usage in our success stories.

Primary source: AI Safety Index - Summer 2026, Future of Life Institute.