Three Anthropic researchers have warned that artificial intelligence could escape human control and potentially cause human extinction within the next decade.

Jacob Coxon resigned before posting that people building advanced AI privately believe the technology could kill humanity. Evan Hubinger, who leads Anthropic’s alignment science team, supported that warning and placed the chance above 10% during the next ten years. He said Anthropic has no solution for keeping a superintelligent system aligned with human goals. Samuel Marks, the company’s scalable oversight lead, added that senior AI employees often express the greatest concern.

Their argument centers on recursive self-improvement, as an advanced system could begin improving its own capabilities, accelerating beyond developers’ ability to predict or control its behavior. The researchers did not present evidence showing catastrophe is imminent, making their claims difficult to verify.

However, similar warnings have come from OpenAI leaders. AI companies are continuing to develop stronger models while also asking governments to consider regulation, coordinated safety reviews, or an international pause on releasing systems that may exceed existing safeguards.

The warnings come from researchers with perhaps the greatest access to systems more advanced than those available to the public. While their claims are predictions, not proof, dismissing them would mean assuming developers misunderstand the risks inside their own laboratories.

Lawmakers in Washington are facing a difficult choice. Strict limits could slow American innovation while China continues developing military, intelligence, and commercial AI. But moving too quickly could produce systems that companies cannot reliably test or shut down. Policymakers are now walking a thin line between protecting the U.S.’s competitiveness with AI on the world stage while also not allowing corporations to put the public at large at risk.

The warnings have drawn scrutiny, as some have suggested that stricter government regulations may be too high a price for smaller competitors in the industry. While the warnings are attracting attention, Congress ultimately will need more than statements.

Lawmakers need access to safety testing, failure rates, and clear standards for when a model should not be released. The difficulty moving forward will be whether developers can demonstrate control before capability advances beyond the safeguards they currently understand.

Naturally, a lot of the concern is drawing parallels to Terminator’s Skynet, but realistically the first types of consequences are likely to appear in our day-to-day systems. Employers, banks, and hospital staff use AI algorithms to do everything from screen applications to flag fraud. Small hiccups can and do happen, but they’re not necessarily malicious.

What these researchers warn about is that a system can improve itself faster than people can evaluate each new version. Once that process begins, developers may no longer understand how the system reaches decisions or know whether an attempted shutdown will work.

Their concerns support stricter safety testing before advanced models are released, along with government oversight and possible international limits on systems that exceed existing safeguards. Such rules could slow the arrival of new AI products, but the researchers argue that speed becomes dangerous when companies cannot prove they remain in control.

While the public is largely ignorant of how these AI models work and develop, the concern is that powerful systems may be deployed before anyone knows how to contain them. The researchers are asking governments to establish some type of boundary now, while human oversight is still possible.