The Pentagon Put ChatGPT and Grok Behind One Secure AI Portal

The Pentagon has added ChatGPT Mil and Grok for Government to GenAI.mil, the secure enterprise portal that launched with Google’s Gemini. Personnel can now choose among several commercial frontier-model families without sending sensitive government work through ordinary consumer accounts.

The portal is important because it treats generative AI as shared infrastructure rather than a collection of personal subscriptions. It also creates a clearer security boundary. The models may feel familiar, but the approved environment, data handling and task authority are what distinguish military deployment from opening a public chatbot.

What each model is being asked to do

ChatGPT Mil focuses on chat, files, projects and custom GPTs. Reporting based on the department’s announcements describes document-heavy, routine unclassified work such as administration, logistics, planning and policy. That is a broad enterprise workload, not an announcement that a chatbot will independently control weapons or make command decisions.

Grok for Government is positioned around reasoning modes, configurable workspaces and reusable playbooks. The Pentagon described uses ranging from market research for acquisition teams to supply-chain management for logisticians. The product is delivered through Starshield AI.

Gemini was the original commercial model inside the portal. Putting multiple providers behind one entry point lets teams compare strengths and gives administrators a place to enforce common access and data policies.

The distinction between capability and authorization matters. A model able to summarize a procurement document is not automatically authorized to approve a purchase. A model able to draft a plan is not the accountable official who signs it.

Why Impact Level 5 matters

The department uses Impact Levels to classify cloud environments. DefenseScoop reports that the approved versions cleared IL5, which covers Controlled Unclassified Information and other sensitive unclassified work.

IL5 is not a classification for secret information. Personnel still need to know what data belongs in the system, what must remain elsewhere and which task-specific restrictions apply. “Secure portal” is not a universal permission slip.

The security improvement is relative to consumer use. An employee should not paste government material into a personal public account with ordinary retention and product settings. A managed portal can set identity, logging, data-use terms and administrative controls for the organization.

The same boundary appears in our guide to rolling out coding agents at work: central access, scoped authority and auditable review matter more than distributing a tool link.

Multi-model does not mean model-neutral

The Pentagon says the architecture is intended to prevent vendor lock-in. Giving users several models can reduce dependence on one provider’s uptime, roadmap and behavior. It also creates operational complexity.

Prompts, file limits, retrieval behavior and refusal policies differ. The same task can receive different answers. If a team treats model choice as a preference menu without evaluation, multi-model access can multiply inconsistency.

A managed deployment needs task routing and evaluation. Routine summarization may be tested across all models. A specialized logistics workflow may use one because it performs better on the actual documents. The decision should be recorded, not inferred from brand familiarity.

The portal should also preserve provenance. An output should identify which model and configuration produced it, what files were available and when the request ran. Otherwise an organization cannot reproduce or audit a decision support artifact.

The user count changes the risk model

GenAI.mil has reportedly onboarded more than 1.7 million unique users from the department’s roughly 3 million personnel. At that scale, a rare failure pattern is no longer rare in aggregate.

Training cannot stop at prompt tips. Users need classification rules, examples of prohibited data, citation expectations, escalation paths and a clear statement of what the system cannot approve. Administrators need anomaly monitoring without turning every use into surveillance of the employee.

The system also needs feedback that distinguishes harmless style complaints from material errors. A wrong meeting summary and a fabricated supply constraint do not carry the same consequence. Severity should drive review and remediation.

Our article on AI eval ownership explains why a score needs a task definition and decision owner. A portal this large needs multiple evals tied to actual mission workflows, not one leaderboard.

Claude’s absence reveals a contract boundary

Anthropic’s Claude is absent after a public dispute over allowable military uses and contractual guardrails. A federal judge recently ruled against the Pentagon’s supply-chain-risk designation, but that legal development does not automatically place Claude inside GenAI.mil.

The episode shows that model procurement is also policy procurement. Providers differ not only in benchmark scores but in their acceptable-use terms and contractual positions. The government decides what access it requires; vendors decide what they will sell; courts can review the legality of government actions.

Teams outside government face a smaller version of the same issue. A product comparison that ignores terms, data retention and permitted uses is incomplete. Model capability cannot override the contract.

What a safe workflow looks like

Start with a bounded task: summarize a known policy, compare two approved documents or extract action items. Require source citations. Keep a human owner for the resulting decision.

Then measure failure. Does the model omit exceptions, invent references or merge outdated and current policy? Test those cases before adding an action that changes a system.

Use least privilege. The model does not need every drive or database because it might be useful someday. Access should follow the current task and user role.

Finally, keep independent records for consequential work. A chat transcript can support reasoning, but the approved plan, order or policy record belongs in the official system of record.

Before expanding the portal, teams can document one approved input set, expected output, reviewer role and escalation path for each workflow. That makes later model comparisons about mission performance instead of preference or familiarity.

What happens next

The Pentagon says features will expand over time. The useful evidence will be adoption by workflow, measured time saved, error patterns and how often users switch models for a justified reason.

GenAI.mil demonstrates that enterprise AI is becoming a control-plane problem. The headline names are ChatGPT, Grok and Gemini. The durable product is the identity, data and evaluation layer around them.

Procurement should also plan for model retirement. Providers rename products, change context limits and deprecate versions. Government records and reusable playbooks need migration tests so an approved workflow does not silently change when its underlying model is replaced. Multi-model access helps only when the portal can preserve task evidence across those transitions.

The portal’s success should therefore be judged by safe completed work, not logins alone. Adoption is an input; verified mission outcomes are the result.

FAQ

What is GenAI.mil? It is the Pentagon’s managed portal for approved commercial generative-AI tools.

Which models are available? Reporting identifies Gemini, ChatGPT Mil and Grok for Government.

Can users put classified information into it? IL5 covers sensitive unclassified data, not a blanket authorization for classified information.

Does the portal make AI outputs official decisions? No. Model access and decision authority remain separate.