OpenAI Agreement With Pentagon Gets Major Update
OpenAI Agreement now says yes to classified Pentagon use of ChatGPT‑style models & no to domestic mass surveillance and autonomous weapons decisions
The OpenAI Agreement Now Bans Domestic Surveillance and Limits Intelligence-Agency Access
After a weekend of backlash, OpenAI and the U.S. Department of Defense have rewritten their agreement.
The revised contract now includes explicit bans on using OpenAI's models for domestic mass surveillance, maintains human responsibility for use of force as a hard requirement, and bars direct use by defense-intelligence agencies like the NSA without a separate deal.
CEO Sam Altman acknowledged that the original agreement "looked opportunistic and sloppy" and was rushed. The amendments clarify boundaries rather than change the underlying cooperation.
The update matters because it sets a visible precedent: a major AI lab insisting on contractual ethical constraints inside classified environments, not just public-facing rhetoric.
What the Original Agreement Allowed
The deal first surfaced on February 27, when OpenAI announced it had reached an agreement with the Pentagon to deploy its models inside a classified DoD network.
The initial contract gave the Pentagon broad latitude. The Department of Defense could use OpenAI's systems for "any lawful applications"—a phrase DoD insisted on to preserve operational flexibility.
OpenAI said it would embed its own safety principles as technical constraints, including prohibitions on domestic mass surveillance and requirements for human responsibility in any use-of-force decisions.
Altman framed the pact as aligning OpenAI's safety principles with U.S. law and policy inside a classified environment.
The timing was politically charged. The announcement came hours after President Trump ordered federal agencies to stop using AI systems from rival Anthropic, escalating an ongoing dispute between Anthropic and the Pentagon over export controls and AI risk.
Why the Original Agreement Triggered Backlash
Within 48 hours, media outlets, civil-liberties groups, and AI-safety researchers raised alarms about the scope of the deal.
Domestic surveillance risk. Critics worried that "any lawful applications" left room for wide-area surveillance, predictive policing, and analysis of domestic communications. This concern intensified because OpenAI had quietly removed an explicit ban on "military and warfare" uses from its public usage policy in early 2024.
Policy drift. OpenAI's 2024 policy rewrite replaced a clear "no military uses" clause with a vaguer ban on using its services "to harm yourself or others." Researchers argued this change weakened the company's public stance against military collaborations.
Opacity around intelligence agencies. Initial coverage did not clearly state whether U.S. defense-intelligence bodies like the NSA or DIA could access OpenAI models under the same umbrella agreement.
The New York Times and other outlets reported that staff inside and outside OpenAI viewed the move as opportunistic, given the company's recent policy softening and the Trump administration's simultaneous break with Anthropic.
What Changed in the Updated Agreement
After the weekend backlash, OpenAI and the Pentagon moved quickly to revise the contract. Reuters confirmed that two major additions now appear in the OpenAI Agreement.
Explicit ban on domestic mass surveillance.
The agreement now specifies that OpenAI's AI systems cannot be used for surveillance of U.S. persons and nationals, with references to relevant federal regulations written into the contract text.
This addresses the loudest concern raised by civil-liberties advocates: that "any lawful applications" could enable mass monitoring of Americans.
No direct use by defense-intelligence agencies.
In a post on X, Altman said the updated OpenAI Agreement clarifies that services will not be used by intelligence agencies within the Department of Defense—such as the NSA—under this contract.
Any future provision of services to those agencies would require a separate, explicitly amended agreement.
Altman told CNBC that the amendments are meant to clarify boundaries, not fundamentally alter the nature of the cooperation.
The Three Red Lines
Beyond the new clauses, the OpenAI Agreement still enforces what the company calls three "red lines" plus a layered protection scheme.
Reuters reported that the red-line principles include:
No domestic mass surveillance. OpenAI's tools cannot be used to conduct large-scale monitoring of U.S. persons. This was always stated as a principle; the updated contract now makes it explicit and legally binding.
Human responsibility for use of force. OpenAI insists that its systems not be used to autonomously make lethal-force decisions or control weapon systems without meaningful human accountability. This means no autonomous weapons—humans must remain in the loop for any use-of-force decision.
Cloud-controlled deployment. Models are deployed via cloud infrastructure that OpenAI operates, not handed over as on-premise weights. This allows OpenAI to enforce its own safety stack and maintain oversight.
Layered Safeguards and Enforcement
OpenAI told Reuters that it maintains control through a broader, multi-layered strategy.
The company retains full control of safety systems and model behavior. Models are accessed via hosted services, with authorized OpenAI personnel in the loop for oversight when needed. Contractual safeguards give OpenAI the ability to cut off or modify usage if DoD actions violate agreed principles.
This structure differs from a traditional defense contract where the vendor hands over software and walks away. OpenAI maintains ongoing operational involvement, which gives it leverage to enforce its stated principles.
At the same time, Reuters notes the Pentagon has signed similar-scale agreements—up to $200 million each—with multiple labs, including Anthropic and Google. The Department of Defense is keen not to let AI creators' safety preferences permanently limit its operational choices.
OpenAI's Shift From "No Military" to Conditional Partnerships
The updated OpenAI Agreement represents the clearest example yet of how OpenAI's military policy has evolved.
In January 2024, OpenAI removed explicit bans on "weapons development" and "military and warfare" from its public usage policy. The company replaced these with a broad rule against using its services to harm people.
Analysts and security experts argued this change opened the door to military and intelligence contracts, even if OpenAI still claimed a high-level "no harm" stance.
The Pentagon deal confirms that OpenAI is now actively supplying models for defense uses—but under contractual constraints around surveillance and autonomous weapons.
OpenAI has moved from a blanket "no military" posture to a more nuanced, contract-based safety regime. The company is comfortable supplying its most advanced models to defense customers as long as it can keep safety and reputational risks bounded by layered controls.
What This Means for the AI-Defense Landscape
The amended OpenAI Agreement is likely to shape how other AI labs approach defense contracts.
It sets a precedent. A major AI lab is insisting on no domestic mass surveillance and human-in-the-loop use of force as contractual obligations, not just blog-post rhetoric. Future deals—whether from OpenAI, Anthropic, Google, or others—will be measured against this standard.
It tests Pentagon tolerance for constraints. The agreement shows that DoD buyers are at least willing to accept vendor-imposed ethical constraints inside classified environments, even if they push for maximum flexibility elsewhere.
It signals a new norm. Carefully scoped defense deals with explicit red lines may become the standard approach for AI labs that want defense revenue without abandoning their public safety commitments.
It confirms OpenAI's strategic direction. The company is not avoiding military work—it's structuring military work in ways that preserve its ability to claim ethical boundaries while still capturing significant government revenue.
The Bottom Line
The OpenAI Agreement with the Pentagon now includes explicit protections that the original deal lacked: no surveillance of U.S. persons, no access by intelligence agencies without a separate contract, and continued requirements for human control over lethal decisions.
Whether these protections hold under operational pressure remains to be seen. The Pentagon's track record suggests it will push boundaries. OpenAI's track record suggests its policies evolve when business opportunities arise.
But for now, the amended agreement represents something new: a major AI company writing its safety principles into a classified defense contract and doing so publicly after being caught rushing a deal that looked careless.
The question isn't whether OpenAI is working with the military. That's settled. The question is whether contract language can actually constrain how the most powerful AI systems get used inside the world's most powerful military—and whether anyone outside the classified environment can verify compliance.