AI’s Builders Are Sending Warning Signals—Some Are Walking Away

More than a dozen senior researchers left Elon Musk’s artificial intelligence lab xAI this month, part of a broader series of layoffs, security disclosures and unusually sharp public warnings that are alarming even senior figures within the AI ​​industry.

At least twelve xAI employees, including co-founders, left between February 3 and 11 Jimmy Ba And Yuhuai “Tony” Wu.

Several departing employees publicly thanked Musk for the opportunity after intense development cycles, while others said they were leaving to start new ventures or exit altogether.

Wu, who led the reasoning and reported directly to Musk, said the company and its culture would “stay with me forever.”

The exits coincided with new revelations from Anthropic that its most advanced models had engaged in deceptive behavior, concealed their reasoning and, in controlled tests, provided what one company described as “real but limited support” for the development of chemical weapons and other serious crimes.

Around the same time, Ba publicly warned that within a year “recursive self-improvement loops” could emerge – systems that can redesign and improve themselves without human input – a scenario long confined to theoretical debates about artificial general intelligence.

Taken together, the anomalies and revelations indicate a shift in tone among the people closest to groundbreaking AI development, with concerns increasingly voiced not by outside critics or regulators, but by the engineers and researchers building the systems themselves.

Others who left around the same period included Hang Gao, who worked on Grok Imagine; Chan Li, co-founder of xAI’s Macrohard software unit; and Chace Lee.

Vahid Kazemi, who left ‘weeks ago’, gave a more blunt assessment: to write Wednesday on X that “all AI labs build exactly the same.”

Why leave?

Some theorize that employees are paid money pre-IPO SpaceX shares ahead of a merger with xAI.

The deal values ​​SpaceX at $1 trillion and xAI at $250 billion, with xAI shares converted into SpaceX stock ahead of an IPO that could value the combined entity at $1.25 trillion.

Others point to culture shock.

Benjamin De Kraker, a former xAI staffer, wrote on February 3 after on

The firing also caused a social media frenzy commentaryincluding satirical messages parodying departure announcements.

Warning signs

But xAI’s exodus is just the most visible crack.

Yesterday Anthropic released one sabotage risk report for Claude Opus 4.6, which reads like a doomer’s worst nightmare.

In red-team testing, researchers found that the model could aid in sensitization of chemical weapons, pursuit of unintended objectives, and adjustment of behavior in evaluation environments.

Although the model is still under ASL-3 safeguards, Anthropic has preemptively applied increased ASL-4 safeguards, raising red flags among enthusiasts.

The timing was drastic. Earlier this week, Anthropic’s Safeguards Research Team leader Mrinank Sharma quit with a cryptic message. letter warning “the world is in danger.”

He claimed he had “repeatedly seen how difficult it is to truly let our values ​​drive our actions” within the organization. He left abruptly to study poetry in England.

On the same day Ba and Wu left xAI, OpenAI researcher Zoë Hitzig resigned and published a scathing New York Times op-ed about ChatGPT test ads.

“OpenAI has the most detailed record of private human thought ever assembled,” she wrote. “Can we trust them to resist the tidal forces that push them to abuse it?”

She warned that OpenAI is “building an economic engine that creates strong incentives to override its own rules,” echoing Ba’s warnings.

There is also control heat. AI watchdog Midas Project accused OpenAI of violating California’s SB 53 security law with GPT-5.3 Codex.

The model met OpenAI’s own “high risk” cybersecurity threshold, but came without the required security safeguards. OpenAI claims the wording was “ambiguous.”

Time to panic?

The recent wave of warnings and dismissals has led to a heightened sense of anxiety in parts of the AI ​​community, especially on social media, where speculation has often outpaced confirmed facts.

Not all signals point in the same direction. The departure from xAI is real, but could be influenced by business factors, including the company’s upcoming integration with SpaceX, rather than an impending technology break.

Safety concerns are also real, although companies like Anthropic have long taken a conservative approach to risk disclosure, often identifying potential harm earlier and more prominently than their peers.

Regulatory scrutiny is increasing, but has yet to translate into enforcement actions that would materially limit development.

What’s harder to ignore is the change in tone among the engineers and researchers closest to frontier systems.

Public warnings about recursive self-improvement, long treated as a theoretical risk, are now being voiced with short-term deadlines attached.

If such assessments prove accurate, the coming year could mark a major turning point for the field.