Altman scrambles to contain fallout from Pentagon deal

Jumping at a government deal your rival just walked away from rarely ends well.

Less than a week after OpenAI announced an expanded defense agreement with the Pentagon, CEO Sam Altman is managing what looks like the company's messiest brand crisis to date. Between protests outside its San Francisco headquarters, subscriber cancellations, and a reported 295% surge in ChatGPT app uninstalls, the internal mood turned so sour that Altman convened an all-hands meeting on Tuesday to quell the storm.

Speaking to employees, Altman called the fallout really painful and acknowledged that rushing the agreement out on a Friday looked opportunistic and sloppy. But he was also surprisingly blunt about the realities of military contracts: when it comes to military operations, the tech lab does not get a vote.

Moving fast into classified networks

The crisis kicked off after rival Anthropic reached an impasse with the Department of Defense. Anthropic had pushed for explicit contractual red lines against using its models for fully autonomous weapons and domestic mass surveillance, balking at the military's demand for unrestricted access across all lawful uses. Following the collapse of those talks, the Pentagon labeled Anthropic a supply-chain risk to national security, and the Trump administration ordered federal agencies to cease using its tech.

OpenAI moved in within 24 hours to secure its own deal, clearing the way to deploy its frontier models across the department's classified networks. While OpenAI claimed its contract still protects against domestic surveillance, Altman made it clear to staff that the military holds the reins.

According to a partial transcript reviewed by CNBC, Altman told employees that Defense Secretary Pete Hegseth retains operational authority. Whether staff think an overseas strike is justified or not, Altman noted, they do not get to weigh in on those decisions. On X, Altman tried to project personal boundaries, writing that he would rather go to jail than obey an unconstitutional order.

Rivals pounce and contracts get rewritten

The blowback forced OpenAI into rapid damage control. Research scientist Noam Brown clarified on social media that OpenAI will not be deploying its models to the NSA or other defense intelligence agencies for now while contractual ambiguities are sorted out.

Meanwhile, competitors wasted no time capitalizing on the controversy. In an internal memo reported by The Information, Anthropic CEO Dario Amodei slammed OpenAI's positioning as safety theater and characterized Altman's messaging as straight up lies, arguing that OpenAI simply wanted to placate workers while Anthropic took real heat to prevent abuses. Anthropic also saw Claude rocket up app store charts in the wake of the public outcry.

Altman, for his part, warned employees about the competitive reality of government work. He argued that if OpenAI’s safety guardrails push the military too far, officials will simply rely on alternatives like Elon Musk's xAI, which has also agreed to deploy models in classified environments and, as Altman put it, will effectively do whatever the government asks.

Why it matters

This episode marks a defining inflection point for the enterprise AI landscape, exposing how thin the line is between national security contracts and catastrophic consumer blowback.

For users and developers, this mess is a sharp reminder that vendor ethics are no longer abstract PR talking points. As core frontier models reach functional parity, brand trust is driving immediate platform migrations. The 295% surge in ChatGPT uninstalls shows that consumer software can face real, measurable consequences when an AI vendor aligns with controversial defense mandates.

For the broader industry, the Pentagon has established a chilling precedent: accept open-ended military utility or risk being officially blacklisted from federal procurement. That dynamic creates an aggressive prisoner's dilemma among frontier labs. When competitors like xAI are willing to provide unconstrained capability on classified networks, any safety stance taken by an individual lab becomes an existential financial risk.

We think OpenAI miscalculated badly here. Trying to seize Anthropic's dropped mantle within 24 hours looked desperate, and Altman's attempt to brand himself as a pragmatic dealmaker fell flat on both fronts. By agreeing to language it had to walk back days later, OpenAI compromised its safety credibility without securing any real operational veto power, handing its fiercest rival an unexpected moral victory.