Multiple U.S. government agencies have raised safety and reliability concerns about xAI’s Grok chatbot, even as the Pentagon has decided to deploy it in classified settings. An internal General Services Administration report concluded that the latest version, Grok-4, does not meet safety and alignment standards for sensitive government operations. The warnings come at a time when federal AI procurement is accelerating rapidly, raising questions about whether cost and speed are outpacing security vetting.
GSA Flags Safety Gaps in Grok-4
A Jan. 15, 2026, executive summary from the GSA directly addressed concerns about Grok, and a longer internal report went further, concluding that Grok-4 falls short of the safety and alignment benchmarks expected for government use. These findings were compiled ahead of a Defense Department decision to adopt the chatbot in classified environments, meaning the Pentagon proceeded despite documented internal objections from fellow federal agencies. According to that internal analysis, Grok-4 struggled with reliably refusing harmful requests, showed uneven performance on fact-sensitive tasks, and lacked the kind of robust auditing tools that evaluators see as prerequisites for deployment in high-risk settings.
The GSA’s assessment carries weight because the agency is the federal government’s primary procurement arm, responsible for vetting tools before they are offered across agencies. When it flags a product as failing alignment standards, that signal typically triggers additional review before deployment, particularly for systems that will handle sensitive national security information. That the Defense Department moved forward anyway suggests a disconnect between procurement oversight and operational adoption, one that could expose classified systems to the exact risks the GSA identified. It also raises basic governance questions: if the central acquisition authority says a model is not ready for sensitive use, but a major department deploys it anyway, the practical power of those safety findings may be more advisory than binding.
A $0.42 Deal That Opened the Door
The friction between the GSA’s safety findings and the Pentagon’s adoption decision traces back to a partnership announced in September 2025. Under the OneGov program, the GSA entered a government-wide agreement with xAI to make Grok available across federal agencies at a cost of $0.42 per agency. That umbrella contract promised to “accelerate” the use of generative AI in government by pre-negotiating terms and centralizing access, allowing agencies to add Grok to their technology stack with minimal paperwork. In practice, it functioned as an on-ramp: once the deal was in place, any participating office could switch on the chatbot with little more than an internal approval.
The price point, $0.42 per agency, is strikingly low and raises a practical question: whether the bargain rate created institutional momentum that outran the security review process. Federal procurement typically involves extensive testing and compliance checks before tools reach classified environments, with agencies expected to follow acquisition rules that embed risk assessments into contract decisions. But the OneGov agreement appears to have created a fast lane, making Grok available for purchase across the government months before the GSA’s own internal team finished evaluating whether the model was safe enough for the job. The result is a procurement pipeline where availability preceded approval, a sequence that inverts the standard order of operations for national security tools and leaves oversight bodies scrambling to catch up.
Pentagon Pushes Ahead Despite Global Backlash
The Defense Department’s decision to embrace Grok has not occurred in a vacuum. The chatbot has already generated significant controversy, including incidents involving manipulated media and antisemitic content that drew global outcry. In those episodes, Grok was linked to outputs that amplified harmful stereotypes and appeared in deceptive contexts, underscoring how large language models can be misused or misfire even when guardrails are in place. For a tool being integrated into classified military operations, the ability to produce misleading or inflammatory material represents a direct operational liability, not just a public relations problem, especially if its outputs are trusted by analysts or commanders under time pressure.
The Pentagon’s willingness to proceed despite these documented issues suggests that the department views Grok’s capabilities as worth the risk, or that internal pressure to adopt AI tools quickly has overridden the caution that safety evaluations are meant to enforce. Reporting on the government’s deliberations indicates that decision-makers were aware of concerns raised in a broader national security review of Grok’s reliability and resilience, yet still opted to test and deploy the model in sensitive contexts. No public statements from xAI or Elon Musk responding specifically to the GSA’s safety conclusions have surfaced in available reporting, leaving an open question about whether the company has addressed or disputed the agency’s technical findings. In the absence of that response, the public record shows a one-sided conversation in which federal evaluators raised alarms and the Defense Department moved forward regardless.
Anthropic Draws a Line the Pentagon Will Not
The contrast with rival AI company Anthropic sharpens the stakes of the Grok situation. Anthropic is currently in a dispute with the Defense Department over AI guardrails, refusing to allow its Claude models to be used in certain military contexts that the company views as incompatible with its safety commitments. According to a recent account of the standoff, Anthropic has resisted pressure to relax restrictions around targeting assistance and other high-risk uses, even when doing so could have secured lucrative defense contracts. Where xAI has made Grok broadly available through a low-cost government-wide deal, Anthropic has taken the opposite approach, drawing boundaries around how its technology can be deployed even when a major government customer is pushing for fewer restrictions.
This divergence reveals a deeper tension in the federal AI market. The government is simultaneously courting multiple AI providers with very different safety philosophies: one company is willing to sell access at minimal cost with limited public guardrails, while another is actively resisting government requests to loosen its controls. The Pentagon’s choice to integrate Grok while Anthropic holds firm on Claude suggests that the path of least resistance, rather than the path of greatest safety, is winning out in practice. For agencies handling classified intelligence, that tradeoff carries consequences that extend well beyond any single chatbot’s performance on a benchmark test. It could also lock the government into relationships with vendors whose risk tolerance does not match the stakes of national security work.
Speed vs. Security in Federal AI Adoption
The core tension running through these developments is structural, not just about one product. The federal government has built a procurement pipeline, through OneGov and related acquisition frameworks, that is optimized for rapid adoption. The $0.42 per agency price tag makes Grok almost frictionless to acquire, especially for small offices that might otherwise struggle to navigate complex contracting processes. Yet the GSA’s own internal review—summarized in the January executive memo—concluded that Grok-4 does not yet meet the safety and alignment standards that agencies typically demand for tools exposed to sensitive or classified data. That mismatch between ease of purchase and readiness for high-risk use is at the heart of the current controversy.
For now, the Pentagon appears committed to pressing ahead with Grok, betting that internal safeguards, compartmentalization, and human oversight can compensate for the shortcomings flagged by civilian evaluators. But as more agencies follow the OneGov pathway into generative AI, the Grok case illustrates how quickly speed can eclipse security in federal technology policy. Unless procurement rules are updated to ensure that safety findings meaningfully constrain deployment decisions, the government risks normalizing a pattern in which powerful AI systems enter classified workflows before they have cleared the very standards designed to keep them in check. In an era when a single flawed output can distort intelligence assessments or decision-making at the highest levels, that is not just a bureaucratic concern. It is a strategic one.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
