Future of Life Institute AI Safety Index — Summer 2026

What it adds beyond the summary

The summary gives the top-line grades and cross-industry findings; the page carries the per-company diagnostics and rankings the reader would actually use to locate a specific vendor. Grading is US GPA (A–F = 4.0–0.0), evidence cutoff June 3, 2026, and grades are averaged expert assessments with individual scores kept confidential.

Rankings and movement

• Anthropic leads five of six domains; OpenAI now leads Risk Assessment on a broader evaluation suite and external testing.
• Meta rose from 6th to 4th; xAI fell from 4th to 7th.
• Three failing grades, one per region: xAI (US), DeepSeek (China), Mistral (Europe). Mistral scored dead last despite the EU's regulatory lead.
• Existential Safety: no company exceeds C-, most D or below. Named efforts (Anthropic constitutional classifiers, OpenAI's call for governance institutions, DeepMind monitoring, Meta loss-of-control provisions) judged "entirely inadequate." Panel questions interpretability and CoT monitorability because "detection is not prevention."

Per-company specifics worth knowing

• Anthropic: criticized for "questionable military engagements," including a reported link to the Minab school strike causing mass civilian deaths; RSP 3.0 walked back pause commitments. Panel wants quantitative/risk-tied thresholds.
• OpenAI: leadership can override the Safety Advisory Group; public-policy positions diverge from stated safety commitments.
• Google DeepMind: unclear which internal body can halt deployment independently of executives; no executive risk officer or independent audit.
• Meta: whistleblower policy undermined by active enforcement of a non-disparagement agreement.
• Z.ai / Alibaba Cloud / DeepSeek: no published safety frameworks; ratings largely reflect the Chinese regulatory environment, not independent leadership. Deferring to government guidance called "complete passivity" on existential safety. Alibaba Cloud and Z.ai deny US allegations of military ties.
• xAI: no evidence of a meaningful safety team; evaluations have "gaping holes (no AI R&D data)"; thresholds effectively non-binding.
• Mistral: leadership "downplays — and at times dismisses — frontier risk."

Cross-cutting: leaders (Anthropic, OpenAI, DeepMind, Meta) weakened unilateral-pause pledges, some with competitor-contingent conditions — panel calls it "moving goalpost." Frameworks published near US/EU deadlines "lack quantitative thresholds, genuinely independent audits, and clear decision authority."

Limits

This is a scored assessment, not a technical resource — no methodology depth beyond the indicator/domain counts (37 indicators, six domains) and the survey approach. Grades are qualitative expert judgments; per-indicator scores and the full framework are on the linked two-page summary/scorecard, not in this text. Findings are frozen at June 3, 2026 and explicitly exclude later events.