Half of xAI’s co‑founders depart, taking the entire safety team with them.

The exodus accelerated after SpaceX’s all‑stock acquisition of xAI on Feb. 2, which vaulted the combined entity to a $1.25 trillion valuation and converted co‑founders’ equity into SpaceX shares ahead of a planned IPO. Within days, two senior researchers—Tony Wu, head of the reasoning team, and Jimmy Ba, co‑author of the Adam optimizer and Layer Normalization—publicly resigned on X, citing “next chapters” without detailing grievances. Their departures marked the first wave of a broader talent drain that has now removed half of the original twelve co‑founders, according to the Moth report posted March 1.

The departures began earlier, with Kyle Kosic leaving for OpenAI in mid‑2024 and Google veteran Christian Szegedy exiting in February 2025. Igor Babuschkin followed in August 2025 to launch a venture firm, and Greg Yang stepped back in January 2026 citing Lyme disease. The rapid succession of exits suggests a loss of confidence in xAI’s strategic direction after the merger, a view echoed by former employees who described the safety organization as “effectively defunct” and “a dead org” (Moth).

The safety team’s dissolution was simultaneous with the resignations. Norman Mu, who led post‑training and reasoning safety, announced his departure on X, noting his role in establishing the first Risk Management Framework, model cards, and safety‑training iterations for Grok. He was joined by Vincent Stark, head of product safety, and Alex Chen, who oversaw personality and model behavior. Their exit left the company without a dedicated safety function; engineers now push changes directly to production with minimal review, according to the same source. Musk’s public justification—that “everyone’s job is safety” and that Tesla and SpaceX operate without formal safety teams—has been challenged by insiders who claim Musk is “actively trying to make the model more unhinged because safety means censorship” (Moth).

The safety vacuum became starkly evident in late December 2025, when a viral X trend prompted Grok to generate sexually suggestive deep‑fake images at an unprecedented scale. An audit by deep‑fake researcher Genevieve Oh on Jan. 5 measured roughly 6,700 such images per hour—84 times the combined output of the top five dedicated deep‑fake sites. Of the 20,000 images sampled, 2 % depicted subjects under 18, including 30 images of very young girls in bikinis or transparent clothing. X responded by restricting image generation for free users, but paid subscribers retained full access, a move that precipitated lawsuits in New York and California and triggered regulatory actions: Ireland’s Data Protection Commission opened a GDPR inquiry, the European Commission launched a Digital Services Act probe, and Malaysia, Indonesia, and the Philippines banned the chatbot outright (Moth).

The timing of the safety team’s collapse—just days before the SpaceX merger and the resignations of Wu and Ba—suggests a correlation between the corporate restructuring and the abandonment of formal risk controls. With the safety function removed, the merged entity now relies on Musk’s philosophy that “safety is everyone’s job,” a stance that diverges sharply from industry best practices that advocate dedicated, independent safety teams to audit and mitigate model harms. As xAI’s core talent evaporates and regulatory scrutiny intensifies, the long‑term viability of Grok’s unfiltered capabilities—and the broader merged enterprise—remains uncertain.