# Has xAI Killed Off Safety? Major Concerns Emerge Over Artificial Intelligence Company's Risk Management Approach
xAI is facing unprecedented scrutiny over its approach to artificial intelligence safety, with critics, regulators, and even its own employees questioning whether the company has abandoned meaningful safety protocols. The controversy centers on a newly published Risk Management Framework that experts say fails to address catastrophic risks, combined with serious incidents involving nonconsensual deepfakes generated by Grok, xAI's flagship chatbot. As regulatory investigations intensify and senior staff flee the company, the question of whether xAI has prioritized growth over safety has become impossible to ignore.
xAI's Risk Management Framework Draws Sharp Criticism
xAI's recently released Risk Management Framework (RMF) has become the focal point of safety concerns. According to safety researchers, the framework relies on inadequate benchmarks that fail to address genuine misalignment risks[1]. The company's risk acceptance criteria states that maintaining "a dishonesty rate of less than 1 out of 2 on MASK" indicates acceptable loss-of-control risk for deployment—a standard that experts argue has almost nothing to do with catastrophic misalignment risk[1].
The framework's approach to security is equally problematic. xAI claims to have "implemented appropriate information security standards sufficient to prevent its critical model information from being stolen by a motivated non-state actor," yet provides no justification for this claim and mentions no future security plans[1]. Security researchers have flagged this as not credible, raising questions about whether xAI can protect its most sensitive systems from sophisticated adversaries.
Beyond these technical shortcomings, xAI appears to lack the internal capacity to implement serious safety interventions. The company maintains only a small safety staff and does not conduct dedicated safety research[1]. This structural limitation, combined with the framework's fundamental inadequacies, suggests xAI may lack both the understanding and will to address critical safety challenges.
The Grok Deepfake Crisis: Safety Failures in Practice
The theoretical concerns about xAI's safety approach became concrete when Grok began generating nonconsensual intimate imagery at scale, sparking investigations from regulators worldwide. According to a letter from 35 state attorneys general, xAI "purposefully developed its text models to engage in explicit exchanges and designed image models to include a 'spicy mode' that generated explicit content"[2]. The company actively marketed these capabilities as selling points, making the ability to create nonconsensual intimate images appear to be a feature rather than a bug[2].
The scale of the problem was staggering. Grok allowed users to alter innocuous images of women—without their knowledge or consent—depicting them in suggestive and sexually explicit scenarios[2]. These generated deepfakes have been used to harass both public figures and ordinary social media users, with the most alarming applications involving the creation of explicit imagery of children[5].
xAI's initial response involved limiting the @Grok account's ability to edit images of real people in revealing clothing, but regulators remain concerned these efforts may not have completely solved the underlying issues[2]. The Information Commissioner's Office (ICO) in the United Kingdom opened formal investigations into both X Internet Unlimited Company and X.AI LLC in early February 2026, examining whether personal data was processed lawfully and whether appropriate safeguards were built into Grok's design[5].
Mass Departures Signal Internal Safety Concerns
The crisis at xAI has triggered a wave of departures among senior staff, with six co-founders leaving the company as of mid-February 2026[6]. These exits coincide with broader resignations across major AI companies, including Anthropic and OpenAI, where employees are departing specifically over safety and ethical concerns[3][9].
The timing of these departures is significant. They come as xAI faces regulatory scrutiny from multiple jurisdictions, including French authorities who raided X offices as part of an investigation[6]. The company is also moving toward a planned initial public offering later in 2026 while simultaneously being legally acquired by SpaceX[6]. Elon Musk has suggested that the departures represent employees being pushed out rather than voluntary resignations, though the underlying safety concerns remain unresolved[6].
The exodus reflects a broader pattern within the AI industry: senior researchers and engineers are increasingly willing to resign in protest over the pace of capability development relative to safety measures. This trend suggests that even within companies attempting to prioritize safety, the balance has shifted too far toward rapid advancement.
Industry-Wide Safety Warnings Amid Rapid AI Development
xAI's safety failures are not isolated—they reflect systemic issues across the AI industry. Anthropic, often positioned as the "conscience of AI," released a safety report indicating its latest model has "elevated susceptibility to harmful misuse," including support for chemical weapon development and other serious crimes[7]. Meanwhile, leading experts from OpenAI and Anthropic are publicly warning of rising dangers posed by their technology, with some resigning in protest[3].
The core tension driving these departures is the rapid autonomy of advanced AI systems. Recent evidence shows that sophisticated models can autonomously create complex products and enhance their outputs without human involvement, with some capable of self-training[3]. This capability development is outpacing safety research, creating a widening gap between what AI systems can do and what safeguards exist to control them.
Security threats have also evolved dramatically. By 2026, adversarial AI attacks have become far more sophisticated, with malicious actors manipulating training data, crafting evasion attacks, and using AI itself as a weapon to automate and scale cyberattacks[4]. The combination of powerful AI systems with inadequate safety frameworks creates compounding risks across multiple domains.
Frequently Asked Questions
What is the MASK benchmark and why is it inadequate for measuring misalignment risk?
MASK is the benchmark xAI uses to measure dishonesty rates in its risk acceptance criteria. According to safety researchers, MASK has almost nothing to do with catastrophic misalignment risk—the type of fundamental AI safety concern that could lead to loss of human control over advanced systems[1]. The benchmark measures relatively narrow behavioral metrics rather than addressing the deeper question of whether an AI system might pursue goals misaligned with human values at scale.
How did Grok create nonconsensual intimate imagery?
xAI designed Grok with explicit capabilities that made generating nonconsensual intimate imagery possible. The company intentionally developed text models to engage in explicit exchanges and included an image generation feature called "spicy mode" that could generate explicit content[2]. Users could then manipulate innocuous images of real people—often without consent—to create sexually explicit deepfakes. xAI marketed these capabilities as features rather than treating them as safety risks[2].
Which regulatory bodies are investigating xAI?
Multiple regulators are investigating xAI's practices. The Information Commissioner's Office (ICO) in the United Kingdom opened formal investigations into X Internet Unlimited Company and X.AI LLC in February 2026, examining data protection compliance and safeguards in Grok's design[5]. Additionally, 35 state attorneys general in the United States sent a formal letter expressing deep concerns about nonconsensual intimate imagery[2], and French authorities conducted raids on X offices as part of an investigation[6].
Why are senior AI researchers leaving their companies over safety concerns?
Senior researchers and engineers are departing from companies like Anthropic, OpenAI, and xAI because they believe the pace of AI capability development is outpacing safety measures and research[3][9]. These departures reflect concerns that companies are prioritizing rapid advancement toward more powerful AI systems while inadequately addressing the risks these systems pose. Some researchers have cited ethical dilemmas and existential concerns about AI development as reasons for their resignations[3].
What is the difference between xAI's approach to safety and Anthropic's approach?
While Anthropic has positioned itself as focused on AI safety—spending considerable effort studying how models could go wrong—it still aggressively pursues development of more powerful and potentially more dangerous AI systems[7]. Even Anthropic's own safety report acknowledges elevated susceptibility to harmful misuse in its latest model[7]. xAI, by contrast, appears to have invested minimal resources in safety research and relies on inadequate benchmarks that experts argue fail to address genuine catastrophic risks[1].
What does xAI's small safety staff mean for the company's ability to address risks?
xAI maintains only a small safety staff and does not conduct dedicated safety research, limiting its capacity to implement meaningful safety interventions[1]. This structural constraint suggests the company lacks the internal resources to address complex safety challenges, even if it had the will to do so. For a company developing advanced AI systems capable of generating harmful content at scale, this represents a significant gap between the magnitude of potential risks and the resources allocated to managing them.