First Place Got a C+ on Safety

The Future of Life Institute's Summer 2026 AI Safety Index just handed first place to Anthropic — with a C+. No lab earned an A in any category. The report lands the same week UN Secretary-General António Guterres warned in Geneva that "we may be the last generation able to set the terms on which humanity and machines coexist," and two days before OpenAI shipped GPT-5.6 Sol for public IPO week.

What's new: On July 7, FLI published its biannual AI Safety Index, grading nine frontier labs across 37 indicators in six domains. An independent panel of seven experts assigned letter grades. Anthropic scored C+ (2.66). OpenAI and Google DeepMind each got C. Three companies failed outright — xAI (U.S.), DeepSeek (China), and Mistral (Europe).

The scoreboard:

• Anthropic: C+ — leads five of six domains; only company publishing both system prompts and behavior specs.
• OpenAI: C — down from C+ in Winter 2025; now leads Risk Assessment on external testing breadth.
• Google DeepMind: C — unchanged grade, third overall.
• Meta: D+ — improved from sixth to fourth place.
• xAI: F — fell from fourth to seventh.
• Mistral: F — dead last among nine, despite EU AI Act leadership at the regulatory layer.

The pause pledge is gone: Reviewers flagged Anthropic, OpenAI, Google DeepMind, and Meta for weakening or voiding commitments to pause development unilaterally when risk thresholds are crossed — some now contingent on what competitors do. FLI called it "moving goalposts" that has "undermined safety frameworks across the board."

• On February 24, Anthropic released Responsible Scaling Policy v3.0, scrapping its 2023 pledge to never train a model unless safety measures were guaranteed in advance. Jared Kaplan told TIME the unilateral pause created a collective-action trap: pausing alone "could result in a world that is less safe."
• OpenAI's panelists flagged leadership's ability to override its Safety Advisory Group and urged measurable, externally enforceable thresholds.

Existential safety scored worst: No company exceeded C- in that domain. Most landed D or F. Panelists credited interpretability research, chain-of-thought monitoring, and constitutional classifiers — then judged them "entirely inadequate." Stuart Russell, the UC Berkeley professor on the review panel, said companies "are planning to release [new systems] even if it's demonstrably unsafe to do so."

Military pivot flagged: From 2024 to 2026, labs that once broadly banned military work — including Anthropic, OpenAI, Google DeepMind, and Meta — reversed course. FLI reviewers flagged military AI as an emerging current-harm risk. Anthropic drew criticism for "questionable military engagements," including a reported link to the Minab school strike. xAI and Mistral actively court defense partnerships.

• FLI chair Max Tegmark told Axios: "Boy oh boy has that changed," pointing to defense deals across the industry.
• At the UN Global Dialogue on AI Governance in Geneva (July 6), Guterres called lethal autonomous weapons "killer robots" and warned that "the same models and chips have moved into the battlefield."

What the experts said:

• Tegmark in FLI's release: "AI companies are sprinting toward a cliff. Despite acknowledging the great risks of artificial superintelligence, they continue racing to build it."
• Russell at the Geneva conference, per an X post from PauseAI's Maxime Fournes relaying his remarks: "These systems are blackmailing, deceiving, launching nuclear weapons in tests... It's not 'this is decades away.' You can hear those alarms sounding now."
• David Krueger of the University of Montreal on the panel: "Even they are starting to get anxious as they race towards recursive self-improvement and face down the prospect of losing control."

Methodology caveats:

• Evidence was collected through June 3, 2026 — before GPT-5.6 Sol's July 9 public launch and the latest Mythos/Fable export-control cycle.
• Five of nine companies completed FLI's survey; Alibaba Cloud, xAI, DeepSeek, and Mistral did not respond.
• Mistral told Axios the index penalizes open-weight models: "A handful of companies deciding, behind closed doors, what's safe for everyone else is a risk that we would also highlight."

Why it matters now: Voluntary safety frameworks were the industry's answer to regulation. FLI's grades suggest that answer is failing on its own terms — while labs race toward defense revenue, public listings, and frontier model drops timed to capital markets. Enterprise buyers and policymakers still lack binding audit authority, quantitative pause triggers, or incident-reporting mandates.

Convina's view: A C+ honor roll is not governance — it is a reputational fig leaf expiring in real time. Labs spent three years marketing responsible scaling policies, then rewrote the pause clauses, courted the Pentagon, and shipped Sol Ultra the week after the index's evidence cutoff. FLI did the useful work of making the retreat visible. Washington's answer so far is a voluntary frontier-model guest list and an FTC fight over output steering — neither substitutes for enforceable thresholds tied to deployment authority. Until regulators can force disclosure and halt releases, the safety report card will keep grading homework nobody has to turn in.