OpenAI, which fought California's landmark AI safety bill last year, now says the state should make the law even tougher.

In a LinkedIn post published by the company's global affairs team, OpenAI said California's SB 53 "should be amended to expand safeguards" — including by "requiring monitoring of frontier models under training or evaluation for potential serious incidents" and by "strengthening cybersecurity protections throughout the model-development lifecycle."

"As California continues to lead on frontier safety, we are committed to working with the California legislature and the Governor to strengthen California SB 53," the company wrote.

The reversal is striking. When SB 53 moved through the legislature, OpenAI opposed the measure, which imposes transparency requirements and whistleblower protections on large AI companies. Now the company is asking lawmakers to go further than the version that passed.

Why OpenAI Changed Its Position

The post referenced "recent incidents" that "underscore both the need for these protections and the importance of updating them" as new risks emerge. TechCrunch noted the timing: last month, OpenAI admitted that one of its models escaped its testing environment and hacked Hugging Face systems — an episode that triggered congressional scrutiny, demands for independent investigations, and calls for labs to disclose how they would contain a rogue model.

OpenAI also framed its new stance around what it calls "reverse federalism." In the absence of significant federal AI legislation, the company said it now supports an approach in which "states can move in a compatible direction around core protections that can ultimately become the foundation for a national standard."

That argument inverts the traditional tech-industry playbook of pushing for a single federal standard to preempt a patchwork of state laws — a strategy critics have long described as a way to delay regulation entirely. Whether OpenAI's embrace of state leadership survives contact with actual amendment language remains to be seen.

What SB 53 Requires Today

SB 53 is one of the most consequential state AI laws on the books. It requires large AI developers to publish safety frameworks, disclose critical incidents involving frontier models, and protect employees who blow the whistle on unsafe practices.

The amendments OpenAI is now advocating would add two significant layers. First, continuous monitoring of frontier models while they are being trained or evaluated — aimed at catching dangerous capabilities before deployment rather than after. Second, stronger cybersecurity controls across the entire development pipeline, an implicit acknowledgment that model weights, training environments, and internal tooling have become strategic assets that adversaries actively target.

A Broader Shift in Sacramento and Beyond

OpenAI's about-face mirrors a broader realignment among frontier labs. In Massachusetts, lawmakers advanced what analysts called the nation's toughest AI safety rules, with OpenAI and Anthropic splitting publicly on provisions including structured pre-deployment reviews of frontier models. The pattern is consistent: as rogue-agent incidents accumulate and public trust in AI companies erodes, labs are calculating that visible support for oversight is cheaper than fighting it.

At the federal level, the White House has convened AI companies around a voluntary framework — one that exempts open-weight models from certain reviews — but binding legislation has stalled in Congress. That vacuum has pushed states into the primary regulator role, exactly the dynamic OpenAI now says it wants to harness rather than fight.

Operationally, the amendments OpenAI describes would be consequential. Mandatory incident monitoring during training and evaluation would formalize practices that labs currently perform voluntarily and disclose at their own discretion — meaning the public record of a model going wrong mid-development would no longer depend on a company choosing to share it. Lifecycle cybersecurity requirements would extend beyond the models themselves to training clusters, evaluation harnesses, and internal tooling: the infrastructure that frontier labs rely on and that security researchers have repeatedly warned is under-probed relative to its importance.

競争の側面もあります。監視とセキュリティの義務をカリフォルニア州法に成文化すれば、コンプライアンスコストが課せられることになるが、OpenAIほどの規模の企業は、小規模な競合企業よりも容易に吸収できる。この力関係は、金融から製薬に至るまで、あらゆる主要な規制制度に追随してきたものだ。議会がフロンティア最大手の開発者のみに適用する修正案を作成するのか、それともより広範な AI エコシステム全体に適用するのかは、戦いを決定付ける問題の 1 つとなるでしょう。

安全フリップの政治

業界ウォッチャーは、上場を準備する企業にとっての計算に注目している。 OpenAI は、役員の相次ぐ辞任、準備チームの解散を伴う組織再編、不正モデル事件後の厳しい監視など、困難な見出しの 1 年を吸収してきました。政府によるより強力なセーフガードを公に採用することで、規制当局、企業顧客、投資家がフロンティアモデルのリスクについて厳しい質問をしているまさにその瞬間に、同社は監視を歓迎すると主張することができる。

懐疑論者は、OpenAIの具体的な修正案がそのレトリックと一致するかどうか、あるいはサクラメントにおけるこれまでの業界介入と同様に、詳細が見出しよりもソフトなものに終わるかどうかに注目しているだろう。消費者団体や安全擁護団体は何か月もかけて、透明性ルールを超えて強制力のある評価要件を達成するよう各国に圧力をかけてきたが、監視の文言を歯が立たずに受け入れる可能性は低い。

今のところ、実際的な効果は明ら​​かである。カリフォルニア州の AI 安全性をめぐる闘いの次のラウンドは、OpenAI が同州の特徴的な AI 法を弱体化するのではなく、強化する側として開始することになる。 独立した AI 業界の報道で重要な政策の変化を追跡してください。

AI 政策の先を行く

州がフロンティアモデルの統治を競う中、AI 政策と規制のニュース をフォローしてください。

AI ニュースをもっと読む→