OpenAI CEO Sam Altman says the company’s newly announced Pentagon deal was “definitely rushed,” and that “the optics don’t look good.” The admission comes as Washington’s fight with Anthropic spills into the open—and as AI labs try to prove they can work with the U.S. government without crossing into domestic surveillance or autonomous weapons.
The backdrop is the collapse of Anthropic’s negotiations with the Pentagon late last week. In the aftermath, President Donald Trump directed federal agencies to stop using Anthropic’s technology after a six-month transition, while Defense Secretary Pete Hegseth said Anthropic would be designated a national security supply-chain risk. Within hours, OpenAI announced it had secured its own agreement to deploy models in classified environments—a move that immediately triggered questions about what, exactly, OpenAI had agreed to and why Anthropic couldn’t.
Weapons of Math Destruction by Cathy O’Neil
A clear look at how high-impact models and scoring systems can produce real-world harm—useful context for the AI surveillance and accountability debate.
Check Price on AmazonThe optics problem isn’t just “AI + military.” It’s the contrast. Anthropic publicly drew red lines around two categories it says it won’t enable: fully autonomous weapons and mass domestic surveillance of Americans. OpenAI says it shares those same principles—but still landed a deal that will place its models inside classified government environments.
To defend the agreement, OpenAI published a blog post outlining what it calls three “red lines” where its models can’t be used: mass domestic surveillance, autonomous weapon systems, and high-stakes automated decisions (including “social credit”-style systems). The company also argues its safeguards go beyond paper promises, describing a “multi-layered” approach that includes controlling the “safety stack,” deploying via cloud, keeping cleared OpenAI personnel in the loop, and relying on contractual and legal protections.
That blog post didn’t settle the argument—it refocused it. Techdirt’s Mike Masnick, for example, argued that contract language referencing compliance with Executive Order 12333 could still allow surveillance in practice, depending on how collection is conducted outside U.S. borders and what data ends up involving U.S. persons. In other words: a “no domestic surveillance” principle can collide with the reality of how intelligence collection is structured.
Army of None by Paul Scharre
A defense-focused primer on autonomous weapons and “human-in-the-loop” constraints—helpful for understanding why AI guardrails are politically explosive.
Check Price on AmazonOpenAI’s response has also leaned heavily on architecture. In a LinkedIn post, OpenAI staffer Felipe Millon argued that deployment architecture matters more than contract language, and that limiting deployment to cloud API can reduce the risk of direct integration into weapons systems or operational hardware. The core pitch: even if people don’t trust clauses, they can trust constraints baked into the delivery model.
Altman has framed the timing as an attempt to cool tensions after the Anthropic standoff turned hostile. But even he has acknowledged the downside: the deal landed fast, created backlash, and arrived at the exact moment users were already reacting to the government’s pressure campaign—helping fuel the weekend shift where Claude overtook ChatGPT in App Store rankings.
For more context on how the Anthropic conflict escalated, see Pentagon vs Anthropic: Claude Standoff Turns Hostile and Anthropic vs Pentagon: Claude AI Use Terms Hit Snag. And for the product-side momentum that’s happening in parallel, there’s Claude Code lands in Slack for dev workflows.