OpenAI is moving deeper into national security work through a new partnership with Anduril Industries, the defense-tech company started by Oculus founder Palmer Luckey in 2017. The companies say they will explore how AI models similar to those behind ChatGPT can help US and allied forces respond to aerial threats.
The agreement centers on drone defense, but it also points to a larger question for the AI industry: how far should systems built for language, images, and reasoning be allowed to support military decisions?
What OpenAI and Anduril plan to build
On Wednesday, Anduril Industries announced a partnership with OpenAI to develop AI models for military use. The stated goal is to help identify and defend against aerial attacks, especially threats from unmanned drones.
The companies are focusing primarily on counter-unmanned aircraft systems, known as CUAS. The announcement also refers to threats from “legacy manned platforms,” meaning crewed aircraft are also part of the defensive picture described in the deal.
Anduril says the AI systems are meant to help humans handle fast-moving information. In its statement, the company said: “As part of the new initiative, Anduril and OpenAI will explore how leading-edge AI models can be leveraged to rapidly synthesize time-sensitive data, reduce the burden on human operators, and improve situational awareness.”
That framing matters. The companies are not describing a system that independently makes lethal decisions. Instead, the current emphasis is on processing incoming data, summarizing what is happening, and helping operators understand a complex environment more quickly.
Why the partnership is a major shift
The agreement is notable because OpenAI once had a clearer public line against weapons development and military warfare. The company still presents itself as a research organization focused on ensuring that artificial general intelligence benefits all of humanity, but its policy posture around military applications has changed over time.
Anduril is not a neutral software vendor. It currently manufactures products that could be used to kill people, including AI-powered assassin drones and rocket motors for missiles. The company says its systems require human operators to make lethal decisions, while also designing products whose autonomous capabilities can be upgraded over time.
OpenAI CEO Sam Altman presented the partnership as a national defense effort. He said: “Our partnership with Anduril will help ensure OpenAI technology protects US military personnel, and will help the national security community understand and responsibly use this technology to keep our citizens safe and free.”
That statement captures the companies’ case for the deal: AI can help protect service members and support national security. Critics of such work, however, are likely to focus on the same point from the opposite direction: if AI helps military operators act faster, it may also move the technology closer to life-or-death decisions.
AI companies are entering defense work
OpenAI is not alone in moving toward defense and government security work. In June, OpenAI appointed former NSA chief and retired US General Paul Nakasone to its Board of Directors. At the time, some experts saw the appointment as a sign that OpenAI might be preparing for more cybersecurity and espionage-related work.
Other major AI companies have also stepped into the sector. Last month, Anthropic partnered with Palantir to process classified government data. Meta has started offering its Llama models to defense partners.
This is a different mood from 2018, when Google employees staged walkouts over military contracts. Now, Google competes with Microsoft and Amazon for Pentagon cloud computing deals. The military market has become difficult for large technology companies to ignore.
The Pentagon has also shown growing interest in AI-powered systems. Its Replicator program aims to deploy thousands of autonomous systems within the next two years. Anduril is helping advance the US military’s interest in drone swarms, according to reporting cited in the source article.
The reliability problem for LLMs
There are already many kinds of artificial intelligence used by the US military. But the systems OpenAI is best known for are large language models, or LLMs, sometimes described as large multimodal models when they work across text, images, and audio.
Those models are trained on massive datasets pulled from many different sources. Their strengths are also the source of concern: they can synthesize information quickly, but they are also known to confabulate erroneous information and can be vulnerable to prompt injections.
That creates a difficult fit for military settings. If an LLM is used to summarize defensive information or assist with target analysis, errors may not be ordinary software mistakes. They could shape decisions under pressure, in environments where speed and accuracy matter at the same time.
The Anduril statement acknowledges the need for guardrails. It says: “Subject to robust oversight, this collaboration will be guided by technically informed protocols emphasizing trust and accountability in the development and employment of advanced AI for national security missions.”
Human control remains the central issue
The central question is not simply whether AI can process more information than a human operator. It is whether the output of that processing can be trusted when the stakes include weapons, aircraft, drones, and battlefield decisions.
For now, the companies describe the role of OpenAI models as support for human decision-making rather than replacement of it. That distinction is important, but it does not end the debate. A system that reduces workload and improves situational awareness can still influence what humans see, prioritize, and decide.
The partnership between OpenAI and Anduril shows how quickly consumer-facing AI technology is moving into national security. It also shows why the industry’s choices around military use are becoming harder to separate from its claims about safety, accountability, and public benefit.