OpenAI and Anthropic researchers decry ‘reckless’ safety culture at Elon Musk’s xAI

AI safety researchers from OpenAi, Anthropic and other organizations publicly speak out against the “reckless” and “fully irresponsible” safety culture at Xai, the AI startup of billion dollars owned by Elon Musk.

The criticism follow weeks of scandals at Xai that have overshadowed the technological progress of the company.

Last week the AI chatbot of the company, grock, anti -Semitic comments and he repeatedly called himself ‘mechhitler’. Shortly after Xai had taken his chatbot offline to tackle the problem, it launched an increasingly capable AI model, Grok 4, which WAN and others found to consult Elon Musk’s personal politics for helping to answer hot button. In the latest development, XAI AI companions launched the form of a hyper-sexualized anime girl and a too aggressive Panda.

Friendly Joshing among employees of competing AI labs is quite normal, but these researchers seem to call for more attention to the safety practices of Xai, who, according to them, are at odds with industrial standards.

“I didn’t want to post on grok safety, because I work at a competitor, but it’s not about competition,” said Boaz Barak, a professor of computer science who is currently on Harvard’s leave to work on safety research at OpenAI, in a Tuesday Post on X. “I appreciate the scientists and engineers at Xai, but the way in which safety was treated is completely irresponsible.”

Barak in particular makes the decision of Xai not to publish system maps – Industrial Standard reports that detailed training methods and safety evaluations in a good confidence in good confidence to share information with the research community. As a result, Barak says that it is unclear which safety training has been done on Grok 4.

OpenAi and Google themselves have a spotty reputation when it comes to the immediate sharing of system cards when revealing new AI models. OpenAi decided not to publish a system card for GPT-4.1 and claimed that it was not a border model. In the meantime, Google was waiting for a security report months after revelation Gemini 2.5 Pro. However, these companies publish historical safety reports for all Frontier AI models before they enter the entire production.

Barrak also notes that the AI companions of Grok “take the worst problems we currently have for emotional dependencies and tries to strengthen them.” In recent years we have seen countless stories by Unstable people who develop with regard to the relationship with chatbotsAnd how AIs can give too high answers about the edge of common sense.

Samuel Marks, an AI safety researcher with Anthropic, also took the decision of Xai not to publish a safety report and to call the move ‘reckless’.

“Anthropic, OpenAi and Google’s release practices have problems,” Marks wrote in one Post on X. “But at least they do something, everything to assess safety before deployment and document findings. Xai does not do that.”

The reality is that we don’t really know what Xai did to test Grok 4. In a broadly shared message in the Online Forum Lesswrong, One anonymous researcher claims that Grok 4 does not have meaningful security money rails Based on their tests.

Whether that is true or not, the world seems to find out in real time about the shortcomings of Grok. Several of Xai’s safety problems have since become viral and the company claims to have tackled them Tweaks on the GROK’s system system.

OpenAi, Anthropic and Xai did not respond to WAN’s request for comment.

Dan Hendrycks, a safety adviser for Xai and director of the Center for AI Safety, Posted on X That the company did “dangerous skills evaluations” on Grok 4, indicating that the company has carried out some pre-deployment for safety problems. However, the results for those evaluations are not publicly shared.

“It concerns me when standard safety practices are not maintained in the AI industry, such as publishing the results of dangerous capacity evaluations,” said Steven Adler, an independent AI researcher who previously led dangerous capacity evaluations at OpenAI, in a statement to WAN. “Governments and the public deserve to know how AI companies deal with the risks of the very powerful systems they say they are building.”

What is interesting about the questionable safety practices of Xai is that Musk has long been one of the most striking proponents of the AI sector. The billionary owner of Xai, Tesla and SpaceX has warned many times about the potential for advanced AI systems to cause catastrophic results for people, and he has praised an open approach for developing AI models.

And yet AI researchers from competing laboratories claim that Xai from the industrial standards are going to safely release AI models. In addition, the startup of Musk unintentionally a strong matter for state and federal legislators can establish rules on publishing AI security reports.

There are various attempts at state level to do this. California State Sen. Scott Wiener pusht a bill for which leading AI LABORATIRATIES – probably including Xai – require that he publishes safety reports, while New York Gov. Kathy Hochul is currently considering a similar bill. Proponents of these accounts note that most AI Laboratories publish this kind of information anyway – but apparently they do not all do it consistently.

Nowadays, AI models should not show real-world scenarios in which they cause really catastrophic damage, such as the death of people or billions of dollars in compensation. Many AI researchers say that this can be a problem in the near future, given the rapid progress of AI models, and the billions of dollars Silicon Valley invests to further improve AI.

But even for skeptics of such catastrophic scenarios, there is a strong case to suggest that the misconduct of Grok contains the products that it contains considerably today.

Grok spread anti -Semitism around the X platform this week, only a few weeks after the chatbot repeatedly raised “white genocide” in conversations with users. Musk has soon indicated that Grok will be more ingrained in Tesla vehicles, and Xai is trying to sell ITS AI models to the Pentagon And other companies. It is difficult to imagine that people who control the cars from Musk, federal employees who protect the US or Enterprise employees who automate tasks will be more receptive to this misconduct than users on X.

Various researchers claim that AI safety and coordination tests not only ensure that the worst results do not take place, but they also protect against behavioral problems in the short term.

Grok’s incidents tend to overshadow the rapid progress of Xai in developing frontier AI models that best OpenAi and Google’s technology, just a few years after the startup was established.