The Pentagon installs ChatGPT and Grok in its secure d

Using a general public artificial intelligence assistant to summarize a sensitive military document would pose an obvious risk. The text could leave the government environment, be kept in the wrong place or be used to improve a commercial product.

The Pentagon now offers another solution with ChatGPT Mil and Grok for Government, launched on August 31, 2026 on GenAI.mil.

These tools are accessible to more than three million U.S. military and civilian employees. They may process certain sensitive information, but they are neither authorized for classified information nor deemed infallible.

Three business models in one portal

GenAI.mil was launched in December 2025 with Google Gemini. The portal is now home to ChatGPT Mil of OpenAI and Grok for Government of Starshield AI.

More than 1.7 million unique users are already connected in nine months, according to the [US Department] (https://www.war.gov/News/Releases/Releases/Article/4586482/departement-of-war-launches-starshield-ais-grok-for-government-on-genamil/).

The presence of several models allows theoretically to choose the tool suitable for each task. One official cited by Defense One describes Gemini as useful for some research, while ChatGPT is more suitable for work on lengthy texts.

This diversity can also reduce reliance on a single supplier. However, it does not make models interchangeable: each has its own capabilities, limitations, safety rules and methods of producing responses.

What can ChatGPT Mil do?

ChatGPT Mil takes over the well-known functions of ChatGPT: conversations, uploading files, projects and custom assistants.

The Pentagon plans to use it for documentary work related to planning, policy, logistics and administration. Examples include reconciliation of maintenance reports, preparation of transportation requirements and planning of disaster response.

The [official release] (https://www.war.gov/News/Releases/Releases/Article/4586352/department-of-war-launches-openais-chatgpt-mil-on-genaimil/) describes a platform capable of accelerating routine tasks. However, it does not provide any independent comparative results to measure the accuracy or time actually saved.

According to OpenAI, data processed in this environment remain separate from its commercial services and are not used to drive its public models. This claim is reported by [Air Force Times] (https://www.airforcetimes.com/industry/techwatch/2026/08/31/the-militarys-chatgpt-is-now-live-via-the-pentagons-genai-platform/).

Grok wants to preserve internal knowledge

Grok for Government offers several modes of reasoning called Auto, Fast and Expert, as well as workspaces, persistent projects and "playbooks".

They must maintain reusable methods and procedures so that the knowledge of a team does not disappear when an employee changes positions.

In particular, the Ministry cites market analysis for procurement and supply chain management. These capabilities are official product descriptions; no audit independent of their reliability in GenAI.mil has been published.

TechCrunch stresses that the simultaneous addition of the two models turns GenAI.mil into a multi-model commercial platform rather than a system based on a military AI fully developed by the government.

What does IL5 certification really mean

ChatGPT Mil and Grok for Government are accredited at impact level 5, or IL5.

This standard allows the processing of Controlled Unclassified Information, often abbreviated CUI. These are non-classified information that nevertheless requires specific protections for access, storage and transmission.

It5 does not mean that all military secrets can be submitted to the model. GenAI.mil is not a platform for classified information.

The Navy's directives also prohibit the entry of information to identify an individual and protected medical information, even if certain other CIU documents are permitted.

Accreditation mainly concerns the computer environment and its controls. It does not certify that each response generated by the AI is accurate, impartial or appropriate.

An AI can produce a very convincing error

A language model can invent a rule, incorrectly summarize a policy or forget an important exception. The assured tone of the answer does not measure its reliability.

A uploaded document may also contain malicious instructions to divert the assistant, an attack known as indirect query injection.

The results should therefore be compared with the original documents, particularly when they relate to expenditure, resource displacement or decisions that may affect people.

Pentagon and OpenAI officials told Defense One that GenAI.mil models are not intended for combat. Separate capabilities are developed for classified networks and military operations.

Prices and availability in Canada

The Pentagon has published no separate cost for ChatGPT Mil or Grok for Government, nor full value of the cloud resources required for their operation.

Costs may vary depending on contracts, computer power, storage, support, security checks, training and number of users.

GenAI.mil is reserved for authorized U.S. Department personnel. No access was announced for the Canadian Armed Forces, the Canadian government or NORAD partners.

Therefore, this platform should not be presented as a service offered in Canada.

Limits or Considerations

The precise versions of the models used behind the names ChatGPT Mil and Grok for Government are not clearly documented publicly.

Nor does the Pentagon publish the detailed results of its cybersecurity trials, error rates, conversation retention rules or the frequency of human intervention.

A choice between several models can reduce commercial dependence, but also complicate audits. A single question could receive three different answers, each based on invisible assumptions.

Finally, a secure portal does not automatically correct a misuse. A document placed in the wrong category or a reply copied without verification can always create an incident.

Conclusion

ChatGPT Mil and Grok for Government bring commercial AI assistants into the daily work of one of the world's largest military organizations.

Their immediate interest mainly concerns paperwork, logistics and document analysis, not the autonomous control of weapons.

GenAI.mil could save time while preventing the use of general public services for certain sensitive information. However, its real results will depend on the quality of responses, the discipline of users and the transparency of post-deployment audits.