The AI Safety Index Just Dropped — and Nobody Passed. That's a Problem for Every No-Code Builder.

The Future of Life Institute published its Summer 2026 AI Safety Index on 7 July. Seven independent experts graded nine of the world's leading AI labs across 37 indicators in six domains: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing.

The highest grade awarded: a C+. Three companies failed outright. No lab earned an A or a B in any single category.

The index's central finding isn't just that AI companies are mediocre at safety. It's that the ones who were supposed to be good at it are actively getting worse. They are walking back commitments, making safety contingent on what competitors do, and racing toward capabilities their own frameworks can't govern.

For no-code builders embedding these models into production apps inside **Bubble** AI Agent workflows, **Zapier** steps, **Make** modules, or any other platform connecting to frontier APIs, this isn't a distant policy debate. It's a structural risk sitting inside your product right now.

## What the index actually found

Here's the scoreboard:

- **Anthropic**: C+ (2.66). Led five of six domains. Best transparency in the group. Also criticised for "questionable military engagements" and quietly weakening its own safety pledges.
- **OpenAI**: C (2.28). Slipped from C+. Leads on risk assessment. Faces a wrongful-death lawsuit alleging loosened safeguards.
- **Google DeepMind**: C (2.01). Settled a Character.ai-linked suicide lawsuit earlier this year, now facing another.
- **Meta**: D+ (1.32). The bright spot: climbed from sixth to fourth. Still not a pass.
- **Z.ai**: D- (0.88), **Alibaba Cloud**: D- (0.87)
- **xAI**, **DeepSeek**, **Mistral**: F. All three failed outright. One from the US, one from China, one from Europe. As Max Tegmark told TIME: "shows this is a global problem."

Existential safety was the weakest domain across the entire industry. No company scored above a C-. Most landed at D or below. The panel called every lab's efforts "entirely inadequate": Anthropic's constitutional classifiers, OpenAI's governance proposals, DeepMind's monitoring commitments, Meta's loss-of-control provisions. All of it. On chain-of-thought monitoring as a safety technique, their verdict was blunt: "detection is not prevention."

And here's the finding that matters most: Anthropic, OpenAI, Google DeepMind, and Meta have all weakened or voided earlier pledges to pause development if their systems approached specific danger thresholds. Some replaced unilateral commitments with competitor-contingent clauses. If one lab races ahead, the others' safety constraints evaporate. Stuart Russell, UC Berkeley professor and panel member, put it plainly: "Companies have backed away from earlier commitments to release new systems only with safety measures appropriate for their capability levels; now, they're planning to release them even if it's demonstrably unsafe to do so."

The voluntary safety system created by AI labs is, in the words of the Axios report, "eroding before governments have put a durable alternative in place."

## What does "C+" actually mean when your app depends on these models?

I think it's worth being precise about what these grades represent, because the letter scale undersells how low the bar is.

The index isn't comparing companies against an ideal. It's comparing them against what seven domain experts consider a minimally adequate safety posture given the capabilities these companies are already deploying. A C+ means Anthropic clears that minimum bar. Barely. Everyone else falls below it.

Now think about what a no-code builder is doing when they wire a frontier model into a production app. You're trusting that:

- The model won't produce harmful outputs in customer-facing contexts
- The API provider has adequate safeguards against misuse
- The provider's own governance will catch problems before they reach your users
- The provider won't degrade its safety posture mid-contract to chase a competitor

The FLI Index says you can't count on any of those things. Not from any vendor. And the people who built their brand on safety are the ones rolling back their commitments fastest.

I've written before about the OWASP security implications of vibe-coded apps: the 5,000-plus apps Red Access found leaking API keys, corporate data, and personal information. This extends that argument. The model providers themselves are the vector.

## Why structured platforms are the safety layer

When you use a structured no-code platform like **Bubble**, **Stacker**, **Webflow**, or **Glide** to embed AI, you're not just calling an API directly. You're operating inside a governance framework the platform enforces by default. That framework gives you things raw model access doesn't.

**Permissions and RBAC.** If a model generates something harmful, who sees it? On a structured platform, permissions are baked into the data layer. A rogue AI output doesn't automatically become a public-facing disaster. It hits the same access controls as everything else.

**Deterministic workflows with AI as a step, not the whole thing.** When a Bubble workflow calls Claude or GPT, the AI is one step in a defined, visual sequence. You can see what happens before and after that call. You can gate it with conditions, route failures, set confidence thresholds. The AI doesn't get to be the whole application; it's a component inside something you control.

**Audit trails.** Structured platforms log what happened, when, and who triggered it. If an AI output causes a problem, you can trace it. For anyone building in regulated industries, that's the difference between "we don't know what happened" and "here's the exact workflow run."

**Visual verification.** You can watch the logic execute. Anyone who's debugged a vibe-coded app where the AI generated 400 lines of invisibly wrong JavaScript knows exactly what I'm talking about.

The point isn't that these platforms magically make AI safe. The point is that they provide the governance layer the model companies themselves are demonstrably failing to provide. And when the model companies are getting C+ grades from independent safety experts, that governance layer stops being optional.

## What builders should do right now

**Pick your model provider like you're doing vendor due diligence.** Anthropic leads on transparency. OpenAI leads on risk assessment. Both have material safety gaps. Meta's Llama models, powering a lot of self-hosted and open-weight deployments, come from a company that scored D+. xAI, DeepSeek, and Mistral failed outright. You might still use their models, but know what you're choosing, especially if you handle client data or regulated information.

**Don't let the AI be the last thing that happens before output reaches a user.** Add a human review step, a content filter, a confidence threshold. Something between the model and your end user that isn't also an AI.

**Use platforms that give you visibility.** If you can't see the workflow, you can't secure it. The visual canvas in Bubble, the logging in Make, the audit trail in Stacker: these aren't UX niceties. They're your safety instrumentation.

**Assume the model provider's safety posture will degrade, not improve, over the course of your contract.** The FLI Index shows commitments being walked back across the board. Build your governance assuming the model's guardrails will weaken. If they strengthen, great. Don't bet on it.

**Push your platform to surface safety scores.** No-code platforms aggregate demand across thousands of builders. If Bubble, Make, or Webflow started showing AI Safety Index scores in their model selector dropdowns, the labs would notice. Platforms that take governance seriously should make vendor safety visible.

## The takeaway

The AI Safety Index is a report card from people who study catastrophic risk for a living, and the message is uncomfortable: the companies building the models you depend on are not operating safely, and they're getting less safe as competition intensifies. That doesn't mean stop using AI. It means stop treating model access as interchangeable and start treating the platform layer — its permissions, workflows, audit trails, and visibility — as your primary safety infrastructure. Because the labs aren't going to provide it for you.