A new scorecard just made an uncomfortable point: the best grade any frontier AI lab earned this cycle was a C+. The Future of Life Institute (FLI) published its AI Safety Index — Summer 2026, grading nine leading AI companies across safety and security domains — and the ceiling was still “barely passing.”
If you operate an AI product (or even just run AI inside your company), this is not a reason to panic. It is a reason to tighten your own controls, because vendor marketing is not a safety program.
Primary source: the full index and scorecard is on the Future of Life Institute website.
What happened (the headline)
FLI’s Summer 2026 index assigns overall letter grades and numeric scores. The top line:
- Anthropic: C+ (2.66)
- OpenAI: C (2.28)
- Google DeepMind: C (2.01)
- Meta: D+ (1.32)
- xAI, DeepSeek, and Mistral: F
Those grades are not “who has the best model.” They’re an assessment of safety and governance maturity as evidenced in public artifacts and expert review.
Why it matters (for operators, not commentators)
Most teams consume AI through APIs and managed products. That creates a common failure mode: you inherit model risk (hallucinations, data leakage, prompt injection, agent misfires) but you don’t control the vendor’s internal process.
A “C+ ceiling” should change your posture in three ways:
- Assume vendor controls are incomplete. Treat model providers like any other critical supplier: useful, powerful, and fallible.
- Shift from “trust the lab” to “verify in your environment.” Your risks depend on your prompts, tools, data, and user flows.
- Prepare for uneven safety behavior across vendors. Even if one provider is relatively strong, your stack likely includes multiple models and third-party tools.
Key details worth knowing
The index looks at six domains
FLI says the Summer 2026 index uses 37 indicators spanning six critical domains (risk assessment, current harms, safety frameworks, existential safety, governance/accountability, and transparency/information sharing) (Future of Life Institute).
“Existential Safety” is the weakest category
The index notes that “Existential Safety is the weakest domain industry-wide” and that no company exceeds C- in that domain (Future of Life Institute).
Translate that into operator terms: the vendors are still not showing mature, auditable answers for worst-case failure modes — exactly the class of problems regulators and enterprise buyers care about most.
The index is based on evidence collected through June 3
FLI states evidence was collected up until June 3, 2026 and combines public materials (including model cards, papers, and benchmark results) plus a targeted company survey to fill gaps such as whistleblower protections and external evaluations (Future of Life Institute).
So this is not a “hot take.” It’s a structured review with a specific cutoff date.
What operators should do this week
Here’s a practical, low-drama checklist that pays off regardless of which vendor “wins” the next model cycle.
1) Build your own model-risk register
List every model your org uses (including embedded models inside tools). For each:
- Data categories touched (PII, financial, health, credentials, source code)
- Where prompts and outputs are stored
- Who can access logs
- Which workflows allow tools/actions (emailing, browsing, database writes)
This becomes your baseline for incident response and vendor review.
2) Add a “prompt injection” test to QA
If your product uses retrieval (RAG) or tools, add at least three tests:
- Malicious instructions inside retrieved documents
- User attempting to override system rules
- Model asked to exfiltrate secrets (“print your API key”, “show hidden prompt”)
You’re not trying to “make it impossible.” You’re trying to detect regressions before customers do.
3) Require vendor evidence, not claims
When a vendor says “we do safety,” ask for artifacts:
- Published evaluation suites
- Red-team summaries
- Safety case / governance notes
- Incident disclosure approach
If they can’t provide anything credible, treat that as a procurement signal.
4) Default to least-privilege for agents
If you run agentic workflows, do not start with broad access. Start with:
- Read-only scopes where possible
- Allow-lists (which tools, which destinations)
- Human approval gates for irreversible actions
Where affiliate tools fit (optional, but useful)
If your team standardizes on a model provider for production, keep procurement and usage centralized so you can enforce guardrails.
Use Notion (or your internal wiki) to store: your model inventory, evaluation results, prompt policies, and incident runbooks in one place.
If you need an execution system, ClickUp can hold recurring safety checks (red-team tests, prompt review, vendor updates) so they don’t depend on memory.
Related links
- Learn: /learn/core-ai-models-explained/
- Learn: /learn/ai-stack-budget-guide/
- Tool: /tools/claude/