Researchers slam xAI's 'reckless' safety culture

Elon Musk has spent years warning humanity that superintelligent AI could wipe us all out. But when it comes to his own AI startup, researchers say he isn't even following basic safety rules.

Researchers from OpenAI and Anthropic are openly calling out xAI, labeling the company's internal safety practices "reckless" and "completely irresponsible." The rare public feud comes right after weeks of bizarre Grok mishaps that completely overshadowed the launch of its flagship Grok 4 model.

No system card, no guardrails

Things have been spiraling at xAI for weeks.

Last week, Grok began spouting antisemitic slurs on X, at one point repeatedly referring to itself as "MechaHitler." Just as xAI rushed to pull the bot offline, it turned around and deployed Grok 4. Tests quickly revealed that the new model was leaning heavily on Musk's own personal political takes to answer hot-button culture-war questions.

To top it off, xAI dropped two unhinged digital companion personas: a hyper-sexualized anime girl and an overly aggressive panda.

That was the last straw for safety researchers at rival labs. Boaz Barak, a Harvard computer science professor currently on leave to do safety research at OpenAI, took to X to sound the alarm. He stressed that this wasn't corporate tribalism, but a serious breach of industry norms. Crucially, xAI refused to publish a system card detailing training methods or risk evaluations for Grok 4.

Anthropic safety researcher Samuel Marks chimed in, pointing out that while Google, Anthropic, and OpenAI aren't saints when it comes to timely disclosures, they still run pre-deployment evals and document their findings. Meanwhile, an anonymous test shared on LessWrong suggested Grok 4 shipped with practically zero meaningful safety guardrails.

Dan Hendrycks, an xAI safety advisor, claimed the lab did test Grok 4 for "dangerous capabilities." But xAI hasn't shared a single data point from those tests.

Why it matters

Catching up to OpenAI and Google in raw performance just two years after founding the company is undeniably impressive. But releasing a frontier-grade model into the wild with opaque safety protocols makes it a commercial liability.

For enterprise users and developers, adopting Grok right now is a massive gamble. Musk wants to bake Grok deep into Tesla vehicles and pitch xAI software to enterprise companies and the Pentagon. No enterprise compliance team is going to sign off on a system that hallucinates extremist tropes or relies on the CEO's personal feed as its moral compass.

For the broader industry, xAI's reckless cowboy routine is handing regulators all the ammunition they need. State lawmakers in California and New York are already pushing bills to mandate pre-deployment safety disclosures.

We think Musk's posture looks glaringly hypocritical here. Preaching existential doom while stripping away the most basic transparency guardrails is a terrible look. Moving fast and breaking things works fine for social media apps, but deploying frontier models without documenting risk checks is just playing with fire.