When Anthropic was founded in 2021 by former OpenAI researchers, the company positioned itself as the conscience of artificial intelligence — a firm built on the principle that AI development should prioritize safety above all else. Four years later, the San Francisco-based company is fielding its Claude AI model for military and intelligence applications, including a reported role in helping the U.S. government build a case against Venezuelan President Nicolás Maduro. The transformation raises pointed questions about whether the AI safety movement can coexist with the demands of national defense contracting.
According to a detailed investigation by NBC News, Anthropic’s technology has been used by U.S. defense and intelligence agencies in ways that go well beyond the sanitized corporate language of “national security partnerships.” The reporting reveals that Claude has been employed to help analyze intelligence related to Venezuela, as the Trump administration has escalated its confrontation with Maduro’s government — a campaign that has included sanctions, diplomatic pressure, and the deployment of U.S. Navy warships to the Caribbean.
From Safety Lab to Defense Contractor
Anthropic’s entry into the defense sector did not happen overnight. The company revised its acceptable use policy in late 2024, quietly removing blanket prohibitions on military and warfare applications of its AI models. The updated policy opened the door for partnerships with defense and intelligence agencies, a shift that Anthropic framed as a responsible engagement with government rather than a retreat from its founding principles. The company has argued that it is better for safety-focused AI firms to work with the U.S. government than to cede that ground to less scrupulous competitors.
The policy change coincided with Anthropic’s deal with Palantir Technologies and Amazon Web Services to make Claude available on Palantir’s defense platform, which holds an Impact Level 6 accreditation — the classification required for handling secret-level national security data. Through this arrangement, U.S. intelligence and defense agencies gained access to one of the most capable large language models in existence, housed within infrastructure already deeply embedded in the Pentagon’s operations. Palantir, led by co-founder Peter Thiel, has long served as a bridge between Silicon Valley and the military-industrial complex.
The Venezuela Connection
The Venezuela dimension of Anthropic’s defense work is particularly striking. As NBC News reported, Claude has been used to process and analyze intelligence related to the Maduro regime as part of the broader U.S. government effort to pressure Venezuela. The Trump administration has taken an increasingly aggressive posture toward Caracas, with officials floating the possibility of military action and the Justice Department pursuing criminal charges against Maduro on narco-terrorism allegations that date back to 2020.
The use of AI in this context illustrates how large language models are being integrated into the intelligence cycle — not as autonomous decision-makers, but as tools to synthesize vast quantities of data, identify patterns, and generate analytical summaries that would take human analysts far longer to produce. Intelligence agencies have long struggled with information overload, and AI models like Claude offer the promise of dramatically accelerating the processing of signals intelligence, open-source reporting, and classified communications.
An Industry-Wide Pivot Toward the Pentagon
Anthropic is far from alone in its turn toward defense work. OpenAI, once equally committed to restricting military applications, removed its prohibition on military use in January 2024 and subsequently struck a deal with the Pentagon. Google has expanded its defense AI contracts after weathering internal protests over Project Maven in 2018. Microsoft has long been a major defense contractor through its Azure Government cloud platform. The pattern is unmistakable: the leading AI companies have collectively decided that the defense market is too large and too strategically important to ignore.
The financial incentives are substantial. The U.S. Department of Defense budget for fiscal year 2025 includes billions earmarked for artificial intelligence and autonomous systems. For AI companies burning through cash at extraordinary rates — Anthropic reportedly spends more than $2 billion annually on compute alone — government contracts offer a stable, high-margin revenue stream that can help justify sky-high private valuations. Anthropic was valued at $61.5 billion in its most recent funding round, according to multiple reports, and demonstrating government revenue makes future fundraising considerably easier.
The Safety Paradox
The tension between Anthropic’s safety mission and its defense ambitions is not merely philosophical — it has practical implications for how the company develops and deploys its technology. Anthropic has built its brand on concepts like “constitutional AI,” a training methodology designed to make models more helpful, harmless, and honest. The company publishes detailed safety research and has called for government regulation of AI systems. Critics argue that this positioning becomes difficult to maintain when the same models are being used to support military operations that could result in kinetic action.
Dario Amodei, Anthropic’s CEO, has addressed this tension publicly, arguing that engagement with democratic governments is consistent with the company’s safety mission. In his view, the greater risk lies in authoritarian regimes — particularly China — developing superior AI capabilities without any safety guardrails. This framing positions defense work not as a contradiction of Anthropic’s values but as an extension of them: by helping the U.S. military and intelligence community, the company is ensuring that the world’s most powerful AI systems remain under democratic oversight.
What the Intelligence Community Actually Gets
The practical applications of Claude within defense and intelligence settings are varied and, in many cases, mundane compared to the dramatic headlines they generate. Large language models are being used to draft reports, translate foreign-language documents, summarize lengthy intelligence assessments, and identify connections across disparate data sets. These are tasks that do not involve autonomous weapons or lethal decision-making, but they nonetheless represent a significant augmentation of the intelligence community’s analytical capacity.
However, the line between analytical support and operational involvement can blur quickly. When an AI model helps synthesize intelligence that informs a military targeting decision or a sanctions designation, the model’s output becomes part of a chain of consequences that extends well beyond a summary document. The Venezuela case is instructive: if Claude’s analysis contributes to intelligence assessments that inform naval deployments or covert operations against the Maduro government, the AI system’s role in the resulting outcomes — however indirect — becomes a matter of legitimate public concern.
Employee Dissent and the Culture War Within
Inside Anthropic, the pivot toward defense work has generated friction. While the company has not experienced the kind of mass employee revolt that hit Google during the Project Maven controversy, there are indications of internal unease. Several AI safety researchers joined Anthropic specifically because of its stated commitment to keeping AI out of military applications. The policy reversal has forced some employees to reconsider their relationship with the company, though public departures over the issue have been limited.
The broader AI safety community has also grappled with Anthropic’s shift. Some researchers and advocates have argued that the company’s credibility on safety issues is undermined by its willingness to serve the defense establishment. Others contend that having safety-conscious engineers involved in military AI development is preferable to leaving the field entirely to defense contractors with less experience in AI alignment and safety testing. The debate mirrors a longstanding divide in the technology industry between engagement and refusal when it comes to controversial government contracts.
The Regulatory Vacuum
One factor enabling Anthropic’s defense expansion is the near-total absence of binding regulation governing the use of AI in military and intelligence contexts. The European Union’s AI Act includes exemptions for national security applications. In the United States, the Trump administration has rolled back Biden-era executive orders on AI safety and signaled a preference for industry self-regulation. Without clear legal frameworks governing how AI models can be used in defense settings, companies like Anthropic are largely left to set their own boundaries.
This regulatory gap means that the most consequential decisions about AI’s role in warfare and intelligence are being made by corporate executives and government procurement officers, not by elected legislators or independent oversight bodies. Anthropic’s acceptable use policy, however thoughtfully crafted, is ultimately a corporate document that can be revised at any time — as the company demonstrated when it removed its military restrictions in 2024.
Where This Leads
The integration of frontier AI models into the U.S. defense and intelligence apparatus is accelerating, and Anthropic’s involvement with Venezuela-related intelligence work represents just one data point in a much larger trend. As AI capabilities continue to advance — with models becoming more capable of reasoning, planning, and processing multimodal data — their utility to military and intelligence agencies will only grow. The question facing the industry, policymakers, and the public is not whether AI will be used in national security contexts, but under what constraints and with what accountability.
For Anthropic, the stakes are particularly high. The company has staked its identity on the proposition that building safe AI and building powerful AI are not mutually exclusive goals. Its defense work will serve as a real-world test of that proposition — one whose results will be measured not in benchmark scores or research papers, but in the consequences of the policies and operations that its technology helps to inform.