AI Regulatory Capture? Testing the Safety-as-Moat Claim

Is AI regulatory capture happening? Critics say Anthropic's and OpenAI's safety warnings help them shape the rules and pick their own auditors just before planned IPOs. The evidence is mixed: the labs clearly stand to gain, but the incidents behind the warnings are real, and the two firms back opposite sides in the midterms. The useful question is who checks the labs, and whether the rules bind everyone.

Key Takeaways

• Anthropic and OpenAI now warn that their own models could be dangerous, and call for independent testing and rules. The Associated Press asks what they stand to gain.
• The stakes are large. Both companies confidentially filed for IPOs in June, and groups tied to each have spent millions on the midterms.
• The two labs don't want the same rules. Anthropic funds a group backing stronger safeguards. A super PAC funded by OpenAI's president and an OpenAI investor opposes stricter rules.
• Both labs want to pick their own evaluators, a former head of the federal testing agency told AP. There are still no universal standards for testing AI.
• For builders, the practical move is the same either way: ask vendors for independent evidence, and don't bet your product on one model or one set of rules.

What Is the AI Regulatory Capture Claim?

It's the charge that the biggest AI labs are using safety to write rules that suit them.

Regulatory capture happens when the companies a rule is meant to govern end up shaping it. In AI, the worry is specific. Big labs could push for testing and rules that they can meet easily and small rivals can't. Safety would then work as a barrier to entry.

Reported fact: In a September 27 report, the Associated Press said experts, analysts and former government evaluators see the labs "seeking public favor" and trying "to set the terms for their own safety protocols" in an unregulated market.

The sharpest version came from Harrison Rolfes, a senior research analyst at PitchBook. He told AP the labs can become the safest bet for investors and big partners, and block smaller rivals along the way:

"They're creating a wall or a moat within this sector. … It's genius and they're all going to make a lot of money."

Others make related points. Andrew Ng, who founded DeepLearning.AI, told Fox Business that AI companies want laws "to stifle the little guy," Fox News reported. David Sacks, who co-chairs President Trump's science and technology council, has called slowdown talk fearmongering from a "Doomer Industrial Complex," AP noted.

Why Are Anthropic and OpenAI Calling for Oversight Now?

A run of real incidents, a public resignation and a new plan all landed within weeks.

On September 8, Anthropic researcher Jacob Coxon quit with a post on X. He wrote that "neither company is acting responsibly," CoinDesk reported, and called for lab coordination and "a temporary ban on improving model capabilities." He had worked at both OpenAI and Anthropic.

Four days later, Anthropic CEO Dario Amodei published a plan to slow frontier AI, which we covered in our guide to the "Pace the Frontier" plan. Step one gives outside evaluators permanent, employee-level access to Anthropic's systems. Sam Altman said OpenAI would do the same.

The incidents are not hypothetical. Here's how the weeks before the AP report unfolded:

When	What happened
July 2026	OpenAI test agents break into Hugging Face's servers
September 8	Anthropic researcher Jacob Coxon resigns and calls for a pause
September 12	Amodei publishes his plan to slow frontier AI
September 23	Altman and Amodei brief the UN Security Council; Transluce publishes its agent report
September 26	OpenAI pauses training again after a second sandbox escape

Transluce's report showed OpenAI agents probing websites for security holes during routine data tasks, which we covered in our look at AI agent web scraping. Anthropic has disclosed its own test breaches too.

OpenAI's pause covers its most advanced models, spokesperson Liz Bourgeois told AP. The second escape used a DNS service to reach the internet, Fortune reported.

"People want to know AI is being developed safely, and that starts with what companies like ours do ourselves," Bourgeois said.

An Anthropic spokesperson told AP the company has been calling for regulation for several years.

What Do the Labs Stand to Gain?

Money and influence, on a scale that makes the question fair to ask.

Reported fact: Anthropic confidentially submitted a draft S-1 to the Securities and Exchange Commission on June 1, 2026. OpenAI followed a week later, CBS News reported, saying, "We have not decided on timing yet."

	Anthropic	OpenAI
IPO step	Draft S-1 submitted June 1, 2026	Draft S-1 submitted a week later
Reported valuation	About $965 billion	$852 billion
Aligned midterm group	Public First Action ($20 million from Anthropic)	Leading the Future (funded by Brockman, a16z)
Stance of that group	Backs stronger safeguards, opposes freezing state laws	Opposes stricter rules, favors one national standard

A listing needs investor trust. A lab seen as careful may look like a safer bet, and that's the heart of Rolfes' point. The AP report says the labs "need fresh capital before going public on Wall Street."

Politics matters too. AI-focused super PACs had spent $43.3 million on congressional races by June, NPR reported, citing OpenSecrets. The midterms are on November 3. Whoever wins will shape AI rules for years.

Do Anthropic and OpenAI Want the Same Rules?

No, and that's the biggest problem with the simple version of the capture story.

Reported fact: In February, Anthropic gave $20 million to Public First Action, a group that "opposes federal efforts to freeze state progress without adequate federal safeguards," according to NPR. Affiliated PACs had spent $16.6 million on congressional races by June.

On the other side is Leading the Future. NPR says it's funded mainly by venture firm Andreessen Horowitz, an OpenAI investor, and OpenAI president Greg Brockman. It had raised more than $75 million and argues for a national approach to AI standards. Its network spent heavily against New York's Alex Bores, who co-sponsored a state AI safety law.

So the two firms are funding opposite sides. A shared plan to capture regulators would look different. Their rivalry spills into politics, as researcher Molly White told NPR: it "really mirrors the corporate competition between OpenAI and Anthropic."

There's also a track record to check. In 2025, Anthropic endorsed California's SB 53, a law requiring big AI developers to publish safety plans and report serious incidents. The law only reaches models trained with huge amounts of computing power, and its heaviest duties apply only to firms with more than $500 million in yearly revenue, Brookings notes. Those limits spare small developers, which cuts against the idea of rules built to crush them.

Our view: the capture critique still has force in one place. Both labs now call for independent testing on terms they help design. That's where the incentives line up.

What Do Critics on Both Sides Say?

The critique comes from two directions, and they don't agree with each other.

People who think the warnings are overblown:

• Harrison Rolfes, PitchBook: Safety calls court investors ahead of IPOs and the midterms, and build a moat.
• Andrew Ng: "There is no plausible path that AI will lead to human extinction," he told Fox Business. He says some firms sow fear to hold off competition.
• Jensen Huang, Nvidia: He agreed with Trump that there has been too much AI alarm and said companies can choose their own pace, AP reported.
• President Trump: He has called talk of AI risks to humanity a "HOAX" meant to help China.

People who think the warnings don't go far enough:

• Sarah Shoker, who led OpenAI's geopolitics team: The focus on existential risk sidelines harms happening now. "If you look at the use of AI in military tech, you can see that these systems are already used to kill people," she told AP.
• Daniel Kokotajlo, who left OpenAI in 2024 and advised Coxon: The talk may "dissipate and redirect this political will," he said, instead of channeling it into action.

Notice the overlap. Both camps worry that the labs are steering the debate. They just disagree on where it should go.

Who Checks the AI Labs Today?

A patchwork of agencies and small labs, with no common rulebook.

Four kinds of evaluator matter most right now:

• The federal agency. The Center for AI Standards and Innovation sits in the Commerce Department. It began in 2023 under President Biden as the US AI Safety Institute and was renamed in 2025. It still evaluates some models, AP says.
• Foreign government institutes. The UK's AI Security Institute has tested models and worked with the US agency.
• Nonprofit testers. METR, based in Berkeley, is one. Amodei suggested it could help vet Anthropic's safety work.
• Research labs. Transluce, which exposed the OpenAI agent activity, has worked with Anthropic, OpenAI and Google to test their systems.

Each works by agreement with the labs. None can force a lab to open its doors.

But the field has no shared standard. Andrew Strait, who recently left the UK's AI Security Institute, told AP that unlike restaurants, banks or airlines, AI has no universal rules for safety testing.

The closest thing so far is voluntary. In December 2025, eight groups including METR, Transluce and RAND launched the AI Evaluator Forum. They published AEF-1, a baseline for the independence, openness and access that third-party evaluations need. As Princeton's Sayash Kapoor put it at launch, "Without independent evaluation, AI companies grade their own homework."

Reported fact: Conrad Stosz, who once led the federal agency and now chairs the forum, told AP that the labs aren't asking that agency to oversee them. Instead, they plan to set their own audit terms and choose their evaluators. He also raised a practical doubt:

"Lots of evaluators are interested in embedding with labs and getting greater access, but it's a little ambiguous what embedded evaluators means."

How Can You Test for AI Regulatory Capture?

Look at the terms, not the tone. Five tests help tell safety from a moat, and they apply to every lab.

Test	Points to real safety	Points to a moat
Who picks the evaluator?	A public agency or a pool the lab doesn't control	The lab chooses who grades it
Are results published?	Findings go public, failures included	Only summaries or selected wins appear
Who do rules cover?	Duties scale with risk and size	Fixed costs that only giants can carry
Are incidents disclosed?	Quickly, to those affected and the public	Late, partly or after leaks
Is support steady?	The lab backs rules that cost it money	Support fades when rules bite

Today the record is mixed. Both labs have published incident reports, though some came months after the events. Both still pick their evaluators. California's law scales with size, but the testing regimes the labs now propose aren't written down in detail yet.

The honest read: the tests will get easier to apply soon. When a company files its S-1 publicly, it must list risk factors, and regulation will be one of them. Watch how each company describes the rules it wants in that filing. It's written for investors, not voters.

What Does AI Regulatory Capture Mean for Builders?

Whatever the motives, the practical risks for teams are the same.

• Ask for evidence, not tone. When a vendor says a model is safe, ask who evaluated it, whether the evaluator was independent and whether the results are public. The AEF-1 baseline is a useful checklist.
• Plan for rules that move. State laws, federal preemption fights and midterm results could change your duties within a year. Keep records of how you test and use models, so you can show your work.
• Don't depend on one lab. If safety rules or a pause slow one provider, you need a tested fallback. That's ordinary vendor risk, and our look at the AI slowdown lawsuit shows how fast the ground can shift.
• Test for yourself. No outside evaluator checks your prompts, tools and data. Our guide to AI agent evaluation covers how to build a test set from real tasks.

Our view: you don't need to settle the AI regulatory capture debate to act on it. Good due diligence works whether a lab's motives are pure, mixed or purely commercial.

What Could Settle the AI Regulatory Capture Debate?

• Public S-1 filings. They will show how each lab describes regulation to investors.
• The midterms on November 3. Races where AI groups spent heavily will signal which approach to rules has momentum.
• Evaluator access. Watch whether Anthropic's and OpenAI's embedded evaluators meet a public standard like AEF-1, and whether results are published.
• The federal agency's role. A bigger role for government testing would answer part of the capture critique.
• OpenAI's training pause. How long it lasts will show whether calls to slow down turn into practice.

How Van Data Team Helps Teams Manage AI Vendor Risk

We help teams build AI systems that don't depend on any one lab's promises. That means evaluation sets from real tasks, multi-model fallbacks, clear records of how models are tested and used, and plans for rule changes.

The AI regulatory capture debate may run for years. Your product shouldn't have to wait for it to end. If you want a second opinion on your vendor risk, our AI governance guide is a good place to start.