When OpenAI announced on February 28 that it had secured a deal allowing the US military to deploy its technologies in classified settings, the timing was as striking as the terms. The agreement came as the Pentagon pushed forward with a newly politicized artificial intelligence strategy during military strikes on Iran, and only after a public clash with rival AI firm Anthropic.
The episode, first detailed by MIT Technology Review, reveals more than a simple contract dispute. It exposes a fundamental divide in Silicon Valley over how far tech companies should go in constraining the use of their tools in warfare and surveillance—and who ultimately decides what is acceptable.
OpenAI CEO Sam Altman acknowledged that negotiations with the Pentagon were “definitely rushed,” following what he described as a public reprimand of Anthropic by defense officials. But OpenAI was quick to frame the deal as principled rather than permissive. In a blog post, the company insisted that its agreement bars the use of its models for autonomous weapons and mass domestic surveillance. Altman emphasized that OpenAI had not accepted the same contract terms Anthropic had rejected.
On the surface, the announcement suggested that OpenAI had achieved a delicate balance—winning a coveted defense contract while preserving its ethical red lines. Yet legal experts and critics argue that the distinction may be more about framing than substance. Where Anthropic reportedly pressed for explicit prohibitions embedded in the contract itself, OpenAI relied on existing law as the primary constraint.
According to Jessica Tillipman, associate dean for government procurement law studies at George Washington University, the contract excerpt shared publicly by OpenAI does not grant the company a standalone right to block government uses it finds objectionable, so long as those uses comply with current law. Instead, it reiterates that the Pentagon must adhere to applicable statutes and policies, including a 2023 Defense Department directive on autonomous weapons and constitutional protections such as the Fourth Amendment.
Critics argue that this legalistic approach offers less assurance than it appears. The Pentagon’s directive on autonomous weapons, for instance, sets guidelines for development and testing but does not ban such systems outright. And while the Fourth Amendment protects against unreasonable searches and seizures, its interpretation has evolved—and been contested—over decades.
Skepticism about relying solely on legal boundaries is fueled in part by history. The surveillance programs exposed by whistleblower Edward Snowden were initially defended as lawful by internal authorities before being ruled unlawful after prolonged legal battles. For those wary of AI-enabled surveillance or autonomous weapons, the memory of those revelations casts doubt on assurances that existing rules are sufficient.
Anthropic’s stance—seeking stronger contractual prohibitions against the use of its AI for autonomous weapons or mass surveillance—earned it support among some OpenAI employees and AI ethics advocates. But it also provoked an extraordinary backlash from the Pentagon. On the eve of US strikes in Tehran, Defense Secretary Pete Hegseth publicly condemned Anthropic, echoing President Donald Trump’s call to halt government cooperation with the company.
Hegseth went further, threatening to classify Anthropic as a supply chain risk and to bar any US military contractor from conducting business with it. Anthropic has said it would challenge such a move in court if pursued. OpenAI, for its part, has publicly opposed the proposed blacklisting, even as it stepped into the vacuum left by Anthropic’s exclusion.
OpenAI maintains that its agreement includes a second safeguard beyond legal compliance: technical controls embedded directly into its models. Boaz Barak, the company’s head of national security partnerships, wrote that OpenAI would not provide the military with versions of its AI stripped of safety features. Instead, the company says it can encode prohibitions—such as bans on mass surveillance or fully autonomous weapons use—into the systems themselves.
Yet significant questions remain about how such safeguards would function in classified environments, particularly under tight timelines. The Pentagon has reportedly granted itself six months to phase out Anthropic’s Claude model from certain classified operations and replace it with systems from OpenAI and xAI. But reports indicate that Claude was still used in operations related to Iran hours after the ban was announced, underscoring the complexity of disentangling AI systems already embedded in military workflows.
The broader issue extends beyond a single contract. Should private technology companies impose moral limits that go beyond the law? Or is it appropriate for elected governments and existing statutes to serve as the ultimate arbiters of acceptable use? The Pentagon’s reaction suggests it sees corporate moral gatekeeping as an unacceptable constraint on national security operations.
For OpenAI, the gamble is multifaceted. Internally, some employees may view the agreement as a pragmatic compromise that ensures influence over how its models are deployed. Others may see it as a retreat from stronger ethical commitments. In an industry where top AI talent is fiercely contested, employee sentiment could prove decisive.
Externally, the arrangement positions OpenAI at the center of a rapidly escalating military AI buildout, with geopolitical tensions in the Middle East serving as a high-stakes testing ground. As strikes on Iran intensify and Washington accelerates its AI strategy, the pressure on companies to align with government priorities is likely to grow.
Whether OpenAI’s approach—anchoring its boundaries in existing law while promising embedded safeguards—will satisfy critics remains uncertain. What is clear is that the company’s compromise represents a pivotal moment in the evolving relationship between Silicon Valley and the Pentagon. As the US military integrates AI more deeply into its operations, the debate over who defines the limits of these systems—engineers, executives, lawmakers, or generals—has moved from theoretical to urgent.
