Pentagon guidance requires checks, traceability for AI-generated code
The Department of War now treats AI-generated code as unverified until it has been reviewed and tested, according to GitLab. DoW Instruction 8430.01 took effect September 8, 2026, and applies to every DoW organization and acquisition program that includes software, regardless of cost.
- DoW Instruction 8430.01 took effect September 8, 2026.
- AI-generated code is treated as unverified until reviewed and tested.
- Teams must track the models, versions and datasets used with AI-assisted code.
- Human approval is required for changes to security-critical and safety-critical functions.
- Approved systems must keep government data and prompts within controlled boundaries.
What DoW Instruction 8430.01 requires
Issued by DoW CIO Kirsten Davies, the instruction sets policy and procedures for modernizing and managing software, with an emphasis on delivering capabilities faster. It directs teams to automate testing and security scanning, reuse existing software and produce evidence as software is built.
The guidance encourages AI-assisted development as a force multiplier for speed and quality, but attaches three conditions. AI’s role must be traceable, its code must be verified so developers remain accountable, and government data must stay within approved boundaries.
That means AI-generated or AI-modified code must meet the same review and testing standard as code written by people. Developers and development teams remain accountable for the security, functionality and integrity of code produced or changed using AI.
Tracking AI involvement and reviewing changes
The traceability requirement calls for records covering which models, versions and datasets were used to write or test code. GitLab says its Dependency Scanning can produce a software bill of materials, while audit events and deployment records are generated during delivery.
Agents in the GitLab Duo Agent Platform act through their own service account, tied to the developer who started them. A commit can therefore identify the agent as acting on behalf of a named person, with the activity recorded in the same audit log as other work. GitLab says security information and event management systems can filter that record as AI activity.
Testing, scans and approval rules run in the same pipeline for changes made by a developer or an agent. Merge request approval policies can require a named reviewer when changes affect protected paths or introduce a vulnerability. Protected branches can also prevent an agent from merging without approval.
GitLab’s Dependency Scanning Auto-Remediation, currently in beta, proposes dependency upgrades and build fixes, but those proposals wait at existing approval gates. The company also says its flows can pause for a human decision before a merge or deployment.
Keeping government data within approved boundaries
The instruction requires approved AI services to ensure that government data and prompts are not used to train external models. Non-public DoW information must remain on approved DoW systems, making model selection and hosting part of the software controls.
GitLab says GitLab Duo Agent Platform Self-Hosted can run on infrastructure controlled by the customer, including air-gapped networks. Customers can choose supported models on premises or in a private cloud, while a self-deployed AI Gateway routes requests within their environment.
For GitLab Dedicated for Government customers, the company describes this AI Gateway arrangement as the standard setup. The announcement positions the platform as a way for defense teams to keep AI-assisted changes governed, auditable and under human control, while applying the same guardrails to AI and non-AI development.
