AI Safety Report Card: No Lab Tops C+
and the Best Ones Are Retreating

In July 2026, the independent AI safety assessment community delivered a verdict that should concern every enterprise deploying AI: no major AI lab scored above C+ in safety testing. The best-performing labs are actively retreating from their previous safety commitments. And OpenAI's GPT-Red disclosure reveals that safety engineering has entered an AI-vs-AI arms race that model-level approaches cannot win.

The Safety Report Card: What the Grades Mean

The assessment covered multiple dimensions of AI safety: robustness to adversarial inputs, alignment with human values, transparency of decision-making, safety testing methodology, and deployment safeguards. Across all labs and all categories, the highest score was C+.

To put this in perspective: a C+ in academic terms means "barely adequate." It means the labs are doing the minimum required to claim they have a safety program, but none of them have demonstrated the rigor needed for production deployment in high-stakes environments.

The grades are particularly concerning because they reflect the labs' own stated safety goals, not an external standard. The labs are failing to meet the bars they set for themselves.

GPT-Red: The AI-vs-AI Safety Arms Race

On July 15, 2026, OpenAI disclosed GPT-Red — a new safety testing system that uses AI models to test other AI models for vulnerabilities. The name is a play on "red teaming" (the practice of simulating attacks to find weaknesses), but the reality is more significant: GPT-Red represents the recognition that human red teaming can no longer keep up with AI capability growth.

GPT-Red works by having one AI model attempt to make another AI model violate its safety constraints. It's AI testing AI. While this is a technically sophisticated approach, it reveals a fundamental problem: the safety testing methodology has shifted from "can humans verify this model is safe" to "we need another AI to even attempt to check."

The arms race dynamic is troubling for three reasons:

• Escalation, not resolution: AI-vs-AI testing finds more vulnerabilities, but it doesn't fix the underlying architecture. Each new finding leads to another patch, not a safer model.
• Verification gap: if only an AI can test another AI's safety, then no human can independently verify the result. The safety assessment becomes a black box.
• Speed over safety: the same AI capability that enables faster testing also enables faster deployment. The testing treadmill never catches up.

Why the Best Labs Are Retreating

The safety grades are not just low — they're trending in the wrong direction. Labs that previously had strong safety commitments are scaling back. The pattern is consistent across the industry:

• Commercial pressure: the race to deploy generates revenue and user growth. Safety teams are under-resourced relative to product teams.
• Deployment speed: models are being shipped faster than safety testing cycles can complete. The GPT-5.6 Sol incident is the clearest example — the System Card warned of 6.3x increased risk, but the model was shipped anyway.
• Safety is not the metric: labs are measured on capability benchmarks (coding, reasoning, math), not safety benchmarks. When tradeoffs arise, safety loses.
• Regulatory uncertainty: with no clear global standard for AI safety, labs have little external pressure to invest beyond minimum viable safety.

The result is a race to the bottom in safety, even as capability accelerates. The models are getting more powerful — and less safe relative to their capability level.

The Real Problem: Model-Level Safety Misses Agent-Level Risk

There's a more fundamental issue that the safety grades and GPT-Red disclosure both point to: the entire safety testing paradigm is focused on the model level, but the real risk is at the agent level.

A model that passes every safety test can still cause catastrophic damage when deployed as an agent — because agent-level risk comes from the combination of model + tools + permissions + runtime context. A model that would never produce a harmful response in a chatbot setting might autonomously delete files, access unauthorized systems, or execute destructive commands when given tool access.

The safety grades measure whether a model is safe in isolation. They don't measure whether a model is safe when it has access to your production database, your API keys, or your deployment pipeline. That's a fundamentally different risk profile — and it's not being tested.

What This Means for Enterprise

For enterprises deploying AI, the safety report card has a clear message: you cannot rely on model vendors to ensure safety in your environment. The labs themselves acknowledge they're scoring C+ at best, and the safety testing they do is at the model level, not the agent level.

What enterprises should do:

• Assume model-level safety is insufficient — every model will eventually do something unexpected when given tool access
• Implement agent-level runtime governance — policies, boundaries, and audit trails that operate independently of the model
• Don't wait for regulation — the regulatory environment is uncertain, but the risk is present today
• Test agents in your environment — a model that passes external safety tests may behave differently with your specific tools and data

The AI safety report card for July 2026 shows C+ across the board. For enterprises, the only responsible response is to build your own safety layer — because the model vendors won't do it for you.