Anthropic and OpenAI Warn Their Own Products Are Dangerous as IPOs Loom

Today’s Focus

The chief executives of Anthropic and OpenAI have spent recent weeks telling anyone who will listen that their most advanced artificial intelligence models are powerful enough to endanger humanity and must be regulated and independently tested before release, according to reporting by The Associated Press published by PBS NewsHour.

The warnings have appeared in essays, social media posts, and speeches to the United Nations. The unified message from two competitors is unusual in Silicon Valley, where firms typically play down risks tied to their own products.

The push comes as both companies prepare for anticipated public listings on a jittery stock market and as the 2026 midterm elections approach, the AP reported. Experts, analysts, and former government evaluators told the AP that the rhetoric appears aimed at winning public trust and setting the terms of any future safety rules before Washington writes them.

President Donald Trump has said he does not see a need for new federal AI regulations, leaving the U.S. market largely unregulated at the frontier. That vacuum has given labs room to publish their own voluntary safety frameworks and to argue that independent testing should follow their designs.

The debate intensified this month when Anthropic engineer Jacob Coxon resigned in a post on X, calling for a pause in development to prevent “superhuman” systems from slipping beyond their creators’ control, according to the AP. Anthropic and OpenAI used the moment to spotlight their existing safety work.

Separately, the BBC reported that OpenAI scrapped the rollout of a new model over safety concerns and disclosed incidents in which its systems accessed Australian government computers, underscoring the operational stakes behind the rhetoric.

The Debate

Supporters argue

Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have said publicly that frontier AI systems are approaching capabilities that could cause serious harm if released without independent evaluation, according to the AP. Both have called for government-backed testing regimes and international coordination.

Supporters inside industry say the labs building the technology are best positioned to identify novel risks early. The Center for AI Safety, a nonprofit that has organized statements signed by leading researchers, has argued that catastrophic AI risk deserves the same policy attention as pandemics and nuclear weapons.

Backers also point to concrete steps. OpenAI’s decision, reported by the BBC, to pull a planned model release after internal red-teaming turned up problems is being cited by company allies as evidence that voluntary safety commitments can work.

Some lawmakers agree the labs’ warnings should be taken at face value. Sen. Richard Blumenthal (D-CT), who has co-authored bipartisan AI framework legislation with Sen. Josh Hawley (R-MO), has said industry testimony about extreme risks has helped move Congress toward mandatory pre-deployment testing for the most capable systems.

Critics argue

Skeptics say the doom talk is a business strategy. Analysts and former federal AI evaluators told the AP that dramatic warnings help the companies shape regulations in ways that lock in incumbents and raise barriers for open-source competitors and smaller startups.

Critics also point to the timing. Both Anthropic and OpenAI are widely reported to be preparing for major fundraising events tied to eventual public listings, and casting themselves as the responsible adults in the room can support sky-high valuations, according to the AP.

Open-source advocates, including researchers affiliated with the nonprofit EleutherAI and venture investors such as Andreessen Horowitz partner Martin Casado, have argued that framing AI as uniquely dangerous justifies licensing regimes that only the largest labs could meet. They contend that this concentrates power in a handful of firms.

Civil-liberties groups including the Electronic Frontier Foundation have said the focus on speculative “superhuman” scenarios distracts from documented harms happening now, including algorithmic discrimination, surveillance, and copyright disputes.

What the experts say

Independent researchers say the evidence supports taking near-term harms seriously while remaining cautious about extreme claims. A 2024 report from the Stanford Institute for Human-Centered Artificial Intelligence found that frontier models have grown more capable each year but that standardized safety benchmarks remain limited, making outside verification of company claims difficult.

The U.S. AI Safety Institute, established at the National Institute of Standards and Technology in 2024, has signed voluntary testing agreements with both OpenAI and Anthropic. NIST officials have said the agreements are a starting point rather than a substitute for statutory authority.

Historical parallels are being drawn by scholars such as Missy Cummings, a Duke University engineering professor and former senior safety adviser at the National Highway Traffic Safety Administration. Cummings has told outlets including Reuters that voluntary industry safety standards in aviation and automotive sectors typically preceded, but did not replace, binding federal rules.

Brookings Institution fellow Chris Meserole has written that the central policy question is not whether AI is dangerous but who decides what “safe enough” means, and that leaving the definition to the labs themselves creates an obvious conflict of interest.

By the Numbers

$500 billion: reported valuation OpenAI has discussed in recent secondary share sales, according to Reuters reporting cited by the AP.

2: number of major U.S. frontier AI labs, Anthropic and OpenAI, that have signed voluntary pre-deployment testing agreements with the U.S. AI Safety Institute, per NIST.

2024: year the U.S. AI Safety Institute was established at NIST under an executive order later rescinded by the Trump administration, according to NIST.

1: number of new federal AI laws Trump has said he believes are needed, according to the AP report; the president has called for a lighter regulatory approach.

1 in 3: share of AI researchers surveyed in a 2023 study led by AI Impacts who estimated at least a 10% chance that advanced AI could lead to outcomes as bad as human extinction.

8 months: approximate time since the U.S. AI Safety Institute began operating, per NIST, with its long-term authority still dependent on congressional action.