A comprehensive safety evaluation of nine major AI companies found that none earned a passing grade, with the highest scorer receiving just a C+ and the most critical safety domain showing near-universal failure across the industry. The Future of Life Institute (FLI), a nonprofit focused on existential risk, released its Summer 2026 AI Safety Index in July, marking the third such evaluation since 2024 and revealing a troubling trend: companies previously seen as safety leaders are walking back their commitments to pause development when approaching dangerous capability levels.

What Grades Did the Nine Companies Actually Receive?

Anthropic topped the field with a score of 2.66 out of 4.0, equivalent to a C+, followed by OpenAI with a C grade (2.28) and Google DeepMind with a C (2.01). The remaining six companies scored lower, with xAI, DeepSeek, and Mistral all receiving failing grades. A panel of seven independent researchers spanning AI alignment, governance, and interpretability expertise conducted the evaluation across 37 different safety indicators organized into six domains.

The scoring methodology relied on publicly available model cards, research papers, benchmark results, and a targeted company survey designed to fill transparency gaps the industry hadn't previously disclosed. Each domain was graded using the standard US GPA scale, with individual reviewer grades kept confidential and final scores calculated as an average across multiple reviewers rather than a single person's subjective judgment.

Why Are Companies Retreating on Safety Commitments?

Perhaps more alarming than the raw scores is a pattern the panel identified directly: Anthropic, OpenAI, Google DeepMind, and Meta, four companies previously viewed as relative leaders in safety practice, have all been observed weakening or walking back earlier commitments to pause development unilaterally when approaching capability red lines. Some have attached competitor-contingent conditions to these pledges, meaning they'll only pause if competitors do the same. The panel calls this phenomenon "moving the goalpost" and writes that it has "undermined safety frameworks across the board".

"Companies previously committed to releasing new systems only with safety measures appropriate for their capability levels, but the current trajectory shows they are planning to release them even if it's demonstrably unsafe to do so," noted Stuart Russell, a UC Berkeley professor whose comment appears in the report.

Stuart Russell, UC Berkeley

This shift represents a fundamental change in how leading AI labs approach the relationship between capability development and safety assurance. Rather than treating safety as a prerequisite for release, companies increasingly appear to be treating it as something that can be addressed after deployment or managed through other means.

Which Safety Domain Is Failing Worst Across the Industry?

Existential Safety emerged as the weakest-performing category across all nine companies evaluated. Not a single company scored above a C-minus in this domain, with most falling to a D or below. This category addresses the most fundamental concern in AI safety: what happens if an AI system becomes so capable that humans lose meaningful control over its actions.

The panel specifically examined several genuine attempts at addressing existential safety risks, including Anthropic's constitutional classifiers, OpenAI's calls for governance institutions, Google DeepMind's monitoring commitments, and Meta's loss-of-control provisions. Despite recognizing these efforts as legitimate attempts, the panel judged them collectively "entirely inadequate" for the scale of the challenge.

How to Interpret These Scores and What They Mean for the Industry

• The Index Functions as External Pressure: Rather than serving as a regulatory tool that directly forces companies to change behavior, the FLI Safety Index currently works as a mechanism for sustained external pressure. Historical data shows some movement: Meta climbed from 6th to 4th place in this edition, while xAI dropped from 4th to 7th, suggesting companies do adjust practices in response to their scores.
• Safety Frameworks Lack Real Enforceability: Several companies have recently published or updated safety frameworks, but these frameworks themselves often lack quantitative thresholds, genuinely independent audits, and clear decision authority. In other words, even as companies release more safety frameworks publicly, most of these frameworks currently "look substantive but lack real enforceability," which is exactly why third-party evaluations like FLI's continue to matter.
• Higher Scores Don't Mean Absence of Problems: Even Anthropic, the highest-scoring company, was explicitly flagged by the panel for the same "moving goalpost" problem affecting the industry. The panel recommended Anthropic "reverse the RSP 3.0 walk-back on pause commitments and restore credibility of commitments," and to "treat prevention as seriously as interpretability and detection," implying its current safety strategy may lean too heavily on after-the-fact detection rather than upfront prevention.

The panel also noted a structural limitation in the report itself: the data collection cut off on June 3, 2026, meaning the evaluation doesn't capture any events or developments that may have occurred since that date. This timing consideration is worth keeping in mind when interpreting the scores, particularly for companies that may have announced new safety initiatives after the evaluation period ended.

How Does the Index Handle Different Regulatory Environments?

The report devotes a dedicated section to China's distinct regulatory context, recognizing that the scoring methodology needed to account for fundamentally different governance structures. In China, national binding instruments such as the Cybersecurity Law and the Generative AI Interim Measures carry direct legal force, with violations risking fines up to 50 million RMB or 5 percent of global revenue. By contrast, the kind of "voluntary commitments" common among US companies mostly correspond in the Chinese context to non-binding recommended standards or promotional local regulations not yet in force, carrying far less weight than national binding rules.

This means the low scores Chinese companies including DeepSeek, Alibaba Cloud, and Z.ai received on indicators like voluntary safety commitments partly reflect a structural difference in regulatory pathways rather than necessarily indicating a lack of safety concern. The panel's company-specific recommendations note that these companies' ratings "largely reflect the Chinese regulatory environment rather than independent safety leadership," meaning the low scores don't yet fully capture these companies' actual safety practices either.

The FLI Safety Index has become one of the most closely watched third-party safety evaluations in the industry since its first release in 2024, with this edition widely cited by major outlets including The New York Times, Financial Times, and TIME. As AI systems continue to grow more capable, the gap between what companies say they're doing for safety and what independent evaluators can verify remains a critical tension shaping the future of AI governance.