Anthropic's 'Safety First' AI Clashes With Pentagon Amid Autonomous Agent Expansion

United States - Ekhbary News Agency

Anthropic's 'Safety First' AI Clashes With Pentagon Amid Autonomous Agent Expansion

Anthropic, a leading artificial intelligence company founded on the principle of "safety first," finds itself at a critical juncture in its relationship with the U.S. Department of Defense (Pentagon). With the recent release of its most sophisticated autonomous agents, the military's demand for these advanced capabilities is intensifying, presenting a profound challenge to the company's core ethical commitments and its strategy for global expansion.

On February 5, Anthropic unveiled Claude Opus 4.6, its most powerful AI model to date. A key feature of this release is its ability to coordinate teams of autonomous agents—multiple AIs that can divide tasks and work in parallel to achieve complex objectives. Just twelve days later, the company introduced Sonnet 4.6, a more cost-effective model that rivals Opus in coding and computer-related skills. While Anthropic's initial computer-controlling models, introduced in late 2024, could barely manage a web browser, Sonnet 4.6 now demonstrates human-level capability in navigating web applications and completing forms, according to the company. Both models boast a working memory capacity large enough to store a small library, enabling more complex and sustained operations.

The enterprise sector now constitutes approximately 80 percent of Anthropic's revenue, and the company recently concluded a funding round of $30 billion, achieving a staggering valuation of $380 billion. By all available metrics, Anthropic stands as one of the fastest-scaling technology companies in history. However, beneath the surface of major product launches and impressive valuations lies a significant threat: the Pentagon has indicated it may designate Anthropic a "supply chain risk"—a label typically reserved for foreign adversaries—unless the company relaxes its restrictions on military use. Such a designation could effectively compel Pentagon contractors to remove Claude from sensitive projects, potentially hindering its adoption in critical defense applications.

Tensions escalated following a January 3 operation where U.S. special operations forces conducted a raid in Venezuela, leading to the capture of Nicolás Maduro. Reports from The Wall Street Journal suggest that Claude was utilized during this operation through Anthropic's partnership with defense contractor Palantir. Axios further reported that this incident exacerbated already strained negotiations regarding the precise permissible uses of Claude. When an Anthropic executive inquired with Palantir about the technology's involvement in the raid, the question reportedly triggered immediate alarms within the Pentagon. While Anthropic has disputed that the outreach was intended to signal disapproval of the specific operation, a senior administration official told Axios that Secretary of Defense Pete Hegseth is "close" to severing ties, adding, "We are going to make sure they pay a price for forcing our hand like this."

This collision highlights a fundamental question: Can a company founded with the mission to prevent AI catastrophe maintain its ethical integrity when its most potent tools—autonomous agents capable of processing vast datasets, identifying patterns, and acting on their conclusions—begin operating within classified military networks? Is a "safety-first" AI truly compatible with a client that demands systems capable of reasoning, planning, and acting autonomously at a military scale?

Anthropic has established two explicit red lines: prohibiting mass surveillance of Americans and banning fully autonomous weapons. CEO Dario Amodei has stated that Anthropic will support "national defense in all ways except those which would make us more like our autocratic adversaries." While other major AI labs, including OpenAI, Google, and xAI, have agreed to relax safeguards for use in the Pentagon's unclassified systems, their tools are not yet deployed within the military's classified networks. The Pentagon, conversely, insists that AI capabilities must be available for "all lawful purposes."

This friction directly tests Anthropic's central thesis. The company was established in 2021 by former OpenAI executives who felt the industry was not adequately prioritizing safety. They positioned Claude as the ethically responsible alternative. In late 2024, Anthropic made Claude accessible on a Palantir platform with a cloud security clearance up to "secret" level, marking Claude, by public accounts, as the first large language model to operate within classified systems.

The current standoff forces a critical examination of whether a "safety-first" identity is sustainable once a technology is embedded in classified military operations and whether such red lines are practically enforceable. Emelia Probasco, a senior fellow at Georgetown's Center for Security and Emerging Technology, notes, "These words seem simple: illegal surveillance of Americans. But when you get down to it, there are whole armies of lawyers who are trying to sort out how to interpret that phrase."

The historical context is also relevant. Following the Edward Snowden revelations, the U.S. government defended the bulk collection of phone metadata—call records detailing who called whom, when, and for how long—arguing that such data did not carry the same privacy protections as the content of conversations. The privacy debate then centered on human analysts examining these records. Now, consider an AI system querying vast datasets, mapping networks, spotting patterns, and flagging individuals of interest. The existing legal framework was designed for an era of human review, not machine-scale analysis.

"In some sense, any kind of mass data collection that you ask an AI to look at is mass surveillance by simple definition," states Peter Asaro, co-founder of the International Committee for Robot Arms Control. Axios reported that a senior official argued there is "considerable gray area" surrounding Anthropic's restrictions and that it is "unworkable for the Pentagon to have to negotiate individual use-cases" with the company. Asaro offers two interpretations of this complaint: a generous one suggesting that defining surveillance is genuinely impossible in the age of AI, and a pessimistic one where "they really want to use those for mass surveillance and autonomous weapons and don't want to say that, so they call it a gray area."

Regarding Anthropic's second red line, autonomous weapons—defined narrowly as systems that select and engage targets without human supervision—the definition might seem manageable. However, Asaro perceives a more troubling gray zone. He points to Israel's reported "Lavender" and "Gospel" systems, which allegedly use AI to generate extensive target lists submitted to human operators for approval before strikes. "You've automated, essentially, the targeting element," Asaro explains, "which is something [that] we're very concerned with and [that is] closely related, even if it falls outside the narrow strict definition." The crucial question is whether Claude, operating within Palantir's systems on classified networks, could be performing similar functions—processing intelligence, identifying patterns, and flagging persons of interest—without Anthropic being able to precisely delineate where analytical work ends and targeting begins.

The Maduro operation serves as a critical test case for this distinction. "If you're collecting data and intelligence to identify targets, but humans are deciding, ‘Okay, this is the list of targets we're actually going to bomb’—then you have that level of human supervision we're trying to require," Asaro says. "On the other hand, you're still becoming reliant on the system for decision-making." This raises concerns about the potential for systems to become quasi-autonomous, even with a human in the loop. Anthropic's challenge lies in ensuring its red lines remain clear and enforceable as its technology becomes integrated into complex, high-stakes military environments where speed and precision are paramount, and the lines between intelligence analysis and targeting are constantly evolving.

Ekhbary News Agency