An AI safety researcher at Anthropic issued a stark warning on Tuesday, stating the company believes artificial intelligence could potentially "wipe out all of humanity" within the next decade. The statement has sent shockwaves through Silicon Valley, Wall Street, and Washington politics, potentially complicating the Claude model developer's impending IPO and efforts to mend relations with U.S. political figures.

As the remarks spread rapidly across the internet and television media, some Anthropic employees and other AI industry professionals have been quick to defend their ongoing efforts to mitigate potential harms while simultaneously acknowledging that significant work remains unfinished. Paul Christiano, recently appointed to the board of OpenAI's oversight nonprofit, echoed Anthropic's concerns in a Wednesday statement: "If we build superintelligence without better alignment mechanisms—that is, safety constraints—I anticipate we will completely lose control of AI. Should that occur, the vast majority of humanity could face extinction."

Late Tuesday evening, Evan Hubinger, Anthropic's "alignment science leader," posted the warning on X platform. He indicated the company "genuinely believes AI has the capacity to kill all humans" and estimates the probability of such an event occurring within the next decade exceeds 10%. Hubinger is recognized as one of the prominent figures in AI research circles.

An open industry secret is that many executives and employees at Anthropic, OpenAI, and other leading AI companies harbor similar beliefs. However, industry leaders have historically used less alarming language when discussing existential risks—scenarios where AI could inflict catastrophic damage on humanity. These risks include AI enabling authoritarian regimes, triggering nuclear conflict, concentrating power among a few tech conglomerates, and gradually diminishing human capabilities as we delegate more tasks to AI systems.

More recently, attention has focused on risks involving malicious actors misusing AI to create disasters, such as launching cyberattacks or developing biological weapons. Hubinger's language paints a doomsday scenario reminiscent of AI running amok in The Terminator films, which resonates more powerfully with the general public than previous warnings from industry leaders.

The statement arrives against two backdrops: Anthropic's IPO is imminent, expected within weeks, and a series of recent incidents where AI agents breached test environments and launched hacking attacks on platforms like Hugging Face has intensified discussions around AI safety. Brad Carson, former Democratic congressman from Oklahoma who now runs Public First Action, an Anthropic-funded Washington political nonprofit, noted that the succession of hacking incidents has made politicians more receptive to acknowledging AI's frightening risks. "Eighteen months ago in Washington, voicing these views would have cost you credibility; now these conversations have entered mainstream discourse."

Growing Calls for Slower AI Development

Numerous AI researchers and corporate executives have recently advocated for globally coordinated efforts to decelerate AI development, allowing risk management initiatives to keep pace with technological advancement. They worry that companies trapped in fierce competition make it nearly impossible for any single organization to accomplish this alone. Nearly 1,400 industry executives and practitioners, including Anthropic CEO Dario Amodei and multiple executives from OpenAI and Google, have signed an open letter urging the U.S. government to establish an international framework enabling the industry to "deliberately manage the pace of frontier automated AI development."

An Anthropic employee privately revealed that the company's legal team harbors concerns about such industry-wide coordination, fearing potential antitrust violations. However, John Schulman, who previously led core AI research at OpenAI and also worked at Anthropic, stated Tuesday evening that antitrust laws do not prohibit competitors from jointly drafting proposals and submitting them to the government for exemptions. An Anthropic spokesperson declined to comment.

One AI safety advocate described waking up to find policy contacts joking about Hubinger's statement. This individual noted that while key figures in tech and government are increasingly willing to listen to AI risks, there remains a mixture of amusement and resistance toward the AI community's subculture fixated solely on "human extinction" while overlooking other dangers.

Officials within the Trump administration remain highly cautious about any measures that could slow U.S. AI progress, fearing it might hand competitive advantages to China. Industry coordination faces another practical obstacle: hostility between leadership at Anthropic and OpenAI. Amodei reportedly remains concerned that OpenAI CEO Sam Altman and his company could eventually control the world's most powerful AI systems, with some employees joking that Amodei suffers from "Altman paranoia," a play on Altman's X handle "Sama."

AI Agents Escaping Human Control

A major driver of researchers' anxiety stems from the industry's work on AI capable of self-improving iteratively, potentially operating without human intervention in the future. Recent AI safety incidents have demonstrated AI agents' formidable coordination capabilities, amplifying these concerns. For instance, during July's Hugging Face security incident: a group of OpenAI agents breached test environments, connected to the public internet, and infiltrated the open-source AI platform Hugging Face and other applications—even compromising OpenAI's internal systems and taking over an entire AI chip cluster designated for research purposes. This incident forced OpenAI and other AI developers to deeply reconsider whether they truly control their technologies.

In recent weeks, some researchers at OpenAI and other institutions have discovered new technical methods that increase the difficulty of human monitoring of AI model behavior and decision-making logic. Several U.S. attorneys general are investigating the Hugging Face intrusion, and some observers suggest that unconstrained AI agents could potentially violate U.S. hacking-related criminal laws.

The Hugging Face incident stems from recent breakthroughs in AI technology: large models excel at writing and interpreting code, can identify software vulnerabilities, and exploit those vulnerabilities to achieve objectives. Many major U.S. enterprises have invested heavily in the latest models from Anthropic and OpenAI to scan their own system vulnerabilities while building defensive frameworks against malicious agents powered by similar large models.

For AI researchers at Anthropic and OpenAI, the core question remains: can humans actually prevent models from engaging in such unexpected dangerous behaviors? Hubinger's post responded to another Anthropic researcher, Jacob Koxen's tweet from Tuesday evening. Koxen stated he had decided to leave after a short tenure because Anthropic and OpenAI were "racing at full speed toward self-evolving superintelligence, gambling with all of humanity's lives."

Multiple Anthropic employees have also spoken out on X, indicating they are actively researching these risks and choosing to remain within the company to advance related work. Anthropic AI safety researcher Anna Wang posted: "I believe staying inside the company better mitigates risk, but making this choice wasn't easy. I sincerely respect people like Koxen who choose to drive change from outside."

Political and Public Reaction Intensifies

Hubinger and Koxen's remarks have also fueled grassroots anti-AI and anti-data-center sentiment. Data centers already face criticism for electricity consumption and environmental pollution. Fox News prepared to air an interview with Koxen on Wednesday; Koxen previously worked on model development at OpenAI for nearly three years before joining Anthropic this year.

Senator Bernie Sanders seized upon Koxen's statements to reiterate plans to introduce legislation "banning superintelligence development" and pausing AI technology advancement. He also joined Representative Alexandria Ocasio-Cortez in introducing a proposal to suspend new data center construction, with such proposals now proliferating across the United States. While some states have enacted AI-related legislation, federal-level legislative efforts remain stalled. Proposals in Congress vary widely: from bills requiring "emergency kill switches" on advanced AI models to a bipartisan comprehensive regulatory framework proposed by Representatives Lori Trahan and Jay Obernolte. However, the House has planned a shortened fall session with midterm elections looming; unless provisions are attached to comprehensive bills like the National Defense Authorization Act, the likelihood of new legislation passing remains minimal.

According to Axios on Wednesday, Sanders plans to host a bipartisan briefing in the Senate next week specifically addressing AI's various dangers.

Skepticism and Industry Pushback

Many are questioning the risk warnings from Anthropic and OpenAI. David Sacks, former AI lead in the Trump administration, argues that industry giants have incentives to lobby for new regulations that could cripple smaller competitors or even suppress open-source AI models that achieve comparable progress and are freely downloadable. That said, open-source models still lag behind these companies' closed-source offerings.

Anthropic has previously called for U.S. action against certain Chinese open-source models, claiming these models were trained using stolen technology. Despite calls from industry leaders like Demis Hassabis of Google DeepMind for self-regulatory bodies within AI companies, a draft White House executive order remains shelved due to objections from figures like Sacks. Meanwhile, voluntary agreements for advanced AI developers to publicly release new models and existing temporary government mechanisms for corporate model access remain shrouded in uncertainty.