Sam Altman admits he messed up. That isn't a sentence you hear often from the man steering the world's most influential AI company, but the fallout from OpenAI’s recent military partnership has left him little choice. By his own admission, the way OpenAI handled its new deal with the Department of Defense (DoD) looked "opportunistic and sloppy."
He's right. It did.
Last week, the Pentagon effectively declared war on Anthropic—OpenAI's primary rival—after CEO Dario Amodei refused to drop safety guardrails concerning mass surveillance and autonomous weapons. Within hours of the Trump administration blacklisting Anthropic as a "supply chain risk," OpenAI stepped into the vacuum, announcing its own classified deal. The timing was more than just awkward; it felt predatory. Now, after a massive public backlash and a reported 295% surge in ChatGPT uninstalls, Altman is scrambling to rewrite the terms.
The rush to fill the Anthropic vacuum
To understand why this looked so bad, you have to look at the timeline. On February 27, 2026, the Pentagon demanded Anthropic allow "all lawful use" of its Claude models. Anthropic held firm on two red lines: no mass domestic surveillance and no fully autonomous lethal weapons. The government responded by treating an American startup like a foreign adversary.
Then came OpenAI.
While Anthropic was being publicly pilloried by Secretary of War Pete Hegseth, Altman was closing a deal. The optics suggested that while one company was willing to die on the hill of AI ethics, the other was happy to pick its pockets. It looked like a "crave and frame" job—caving to the Pentagon's demands while framing it as a strategic victory for "national de-escalation."
Amending the "sloppy" language
On Monday, Altman took to X (formerly Twitter) to announce that OpenAI is already modifying the contract. The goal is to make their "principles" clearer, but the shift reveals just how vague the initial agreement really was.
The new language explicitly states that OpenAI’s systems "shall not be intentionally used for domestic surveillance of US persons and nationals." It specifically calls out the Fourth Amendment and the Foreign Intelligence Surveillance Act (FISA). More importantly, it now bars the National Security Agency (NSA) from using these tools without a separate, follow-on contract.
What the updated contract says
- No Domestic Spying: Prohibits deliberate tracking or monitoring of Americans, including the use of commercially bought data like location history.
- Intelligence Agency Ban: Explicitly excludes the NSA and other Department of War intelligence arms from current access.
- Cloud-Only Deployment: By keeping the AI on OpenAI’s cloud instead of "edge" devices, the company claims it can technically prevent the models from being integrated into hardware for autonomous killing.
Altman even went as far as saying he'd rather go to jail than follow an unconstitutional order. It's a bold statement, but it highlights the desperation to win back a user base that is currently flocking to Anthropic.
Why the "safety stack" might be a facade
Despite the new edits, skeptics aren't buying it. The core of the problem is the phrase "any lawful use." In the US, many forms of bulk data collection are technically legal under current interpretations of national security laws. If the Pentagon uses GPT-5 to analyze metadata that the government has "lawfully" acquired, is that a violation of OpenAI's terms?
Critics like Sarah Shoker and researchers at UC Berkeley argue that terms like "unconstrained monitoring" are dangerously vague. What OpenAI calls a "multi-layered safety stack," others see as a collection of escape hatches. If the government decides a specific mission requires "lawful" surveillance, OpenAI’s contractual "red lines" might offer about as much protection as a paper umbrella in a hurricane.
The cost of moving too fast
The market reaction has been brutal. Data from Sensor Tower shows that while ChatGPT uninstalls spiked, Anthropic’s Claude app hit the number one spot on the US App Store. Users are voting with their thumbs. They’re choosing the company that got kicked out of the Pentagon over the one that rushed in to replace them.
Inside OpenAI, the mood isn't much better. Research scientists like Aiden McLaughlin have publicly questioned if the deal was worth the reputational damage. When your own safety researchers are "overwhelmed" by the ethical implications of your business deals, you have a culture problem, not just a PR problem.
What you should do now
If you’re a developer or a business owner relying on these models, the landscape has shifted. Reliability isn't just about uptime anymore; it's about the ethical stability of your provider.
- Audit your AI stack: If your brand depends on "ethical AI," having your primary vendor tied up in a domestic surveillance controversy is a liability.
- Diversify your LLM usage: Don't put all your eggs in the OpenAI basket. If you haven't explored Claude 3.5 or 4, now is the time to set up that API.
- Watch the "Supply Chain Risk" litigation: Anthropic is expected to challenge its designation in court. The outcome of that case will determine whether the US government can legally force AI labs to remove safety guardrails.
OpenAI tried to play the role of the adult in the room, attempting to "de-escalate" the tension between Silicon Valley and the Trump administration. Instead, they’ve ended up looking like they prioritize government contracts over user trust. Altman’s "learning experience" might be a very expensive one.
Keep a close eye on the all-hands meeting scheduled for Wednesday. That's when we'll see if the rank-and-file at OpenAI are actually buying what Altman is selling.
