# Frontier AI Safety: Should the Most Powerful Models Be Subject to Different Rules?

Author: Norbert Biedrzycki
Published: 2025-10-15
Language: en
URL: https://norbertbiedrzycki.pl/en/frontier-ai-safety-should-the-most-powerful-models-be-subject-to/
Summary: At the end of September, California's governor signed SB 53, the Transparency in Frontier Artificial Intelligence Act. It requires developers of the most capable models to publish their safety…

At the end of September, California's governor signed SB 53, the Transparency in Frontier Artificial Intelligence Act. It requires developers of the most capable models to publish their safety frameworks, report critical safety incidents and protect whistleblowers. A year earlier a more ambitious bill, SB 1047, had been vetoed. The narrower version passed.

In July, the European Commission published the General Purpose AI Code of Practice, a voluntary instrument to help providers comply with the AI Act's obligations for general purpose models, including a dedicated chapter on safety and security for models with systemic risk. Most leading developers signed it. Some did not, or signed only parts of it.

In January, the first International AI Safety Report, led by Yoshua Bengio and written by experts nominated by thirty countries, summarized what is known about the risks of general purpose AI. Its main message was uncertainty: capabilities are advancing quickly, and our ability to measure and predict them lags behind.

## Why frontier models are treated differently

Most AI regulation focuses on uses. A credit model is regulated because credit decisions matter. Frontier safety introduces a different logic: some risks come from the capabilities of the model itself, regardless of how a particular customer uses it.

The concerns are specific. Models that provide meaningful help to someone trying to build biological or chemical weapons. Models that significantly increase the capability of cyber attackers. Models that act autonomously in ways that are hard to supervise or correct. And the broader possibility that increasingly capable systems behave in ways their developers did not anticipate.

The leading developers have responded with their own frameworks. Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework all define capability thresholds and the safeguards that must be in place before models crossing them are deployed.

## The credibility problem

These frameworks are a genuine step forward. They also share a structural weakness. In most cases, the organization developing the model also designs the tests, runs the evaluations and decides whether the thresholds have been crossed. The companies are competing intensely, under enormous commercial pressure, in a market where being a few months late can matter.

In any other industry with potentially severe risks, we would not accept self certification as the final word. Pharmaceuticals, aviation and nuclear energy all rely on independent assessment. AI safety institutes in the UK and US, and the EU's AI Office, are early attempts to build that capacity. They are still small relative to the industry they observe.

## Why investors and boards should care

For investors, frontier safety introduces a risk that does not fit neatly into traditional models: capabilities may develop faster than the mechanisms designed to control them. A serious incident involving a frontier model could trigger rapid regulation that changes the economics of the entire sector, including companies building on top of those models.

For enterprises, the question is more immediate. Which models are we building on, what do we know about how they were tested, and how dependent are we on a provider whose safety practices we cannot verify?

> *Leadership question: Should the most powerful AI models be subject to independent safety assessments before they are made widely available?*
