GSA LLM data safeguarding clause (2026)

Law & policy information

Name
Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems

Jurisdiction
United States — federal; GSA acquisitions

Instrument type
Acquisition-regulation class deviation and contract clause

Bill / legal citation
RGO-2026-01; Part 539.71; clause 552.239-7001 (September 2026).

Legislative or adoption status
Part 552 implementation date reached October 19, 2026; application depends on acquisition scope and contract implementation.

Effective dates
Attachment A: Part 552 October 19, 2026; Part 539 March 16, 2026. Earlier use permitted; existing-contract implementation is separately addressed.

Last source verification
October 19, 2026 — fresh PDF matched the previously reviewed memorandum, date table, relevant Part 539 and complete LLM clause.

The GSA LLM data safeguarding clause is clause 552.239-7001, Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems, in the General Services Administration’s RGO-2026-01 acquisition-regulation class deviations. The September 2026 clause addresses government data, contractor oversight, intellectual property, transparency, security incidents and model changes when GSA procures covered LLM functionality.

Status and implementation

The Part 552 deviation, which contains the clause, reached its Attachment A effective date on October 19, 2026. It separately lists March 16, 2026 for Part 539. The table permits earlier use and requires GSA contracting activities to begin using the applicable deviated language no later than its listed date. These part-level dates should be distinguished from the September 2026 version label on the LLM clause.

The memorandum applies to GSA contracting activities and procurements and directs use of the deviation text in place of corresponding GSAR provisions. It distinguishes new solicitations and contracts from open solicitations and awarded contracts, for which contracting officers have implementation discretion. Existing Federal Supply Schedule contracts follow standard refresh and mass-modification protocols. The date table does not itself amend every existing contract. The memorandum describes subsequent formal rulemaking with public notice and comment.

Covered acquisitions and exclusions

Part 539.71 prescribes the clause for procurements of LLMs, generative assistants, chatbots, agentic systems and similar tools where LLM functionality is material and government data will be submitted directly to or produced by the LLM. It includes commercial products and services. Exceptions address internal contractor support systems not delivered to or accessed by government, and incidental or ancillary LLM functionality under the stated conditions. The clause also contains corresponding self-deletion provisions unless the contracting officer identifies otherwise.

Applicable requirements flow down to subcontractors performing specified AI design, development, deployment, operation or monitoring functions involving government data. Certain origin, jurisdiction and foreign-control requirements do not flow down to qualifying open components or fully open models. The clause distinguishes fully open models from models with publicly available weights alone.

Data and intellectual-property protections

Government data comprises defined inputs and outputs, with exclusions for specified background intellectual property and technical system-level information. The government retains its data and custom developments; contractors retain underlying LLMs, base models and background data. Permitted uses are tied to contract performance, tightly bounded security telemetry, support and other written contracting-officer authorization. Prohibited examples include unauthorized model training, commercialization and sale of government data.

Safeguards address access, processing, agreed locations, logical segregation, incident records, deletion and government-configurable privacy tools. The clause recognizes FedRAMP equivalence or overlapping submissions in specified circumstances. It also requires data export and migration support and bars proprietary formats, technical restrictions, added costs or licensing conditions that materially impair government retrieval, use or migration.

Oversight, reporting and remedies

The clause requires notification within 72 hours of actual knowledge of a material clause violation. A separate security-incident provision requires notification as soon as possible and within 72 hours of discovery, daily updates until resolution and preservation of relevant incident artifacts for at least 90 days, subject to its stated reporting provisions.

Other provisions address LLM disclosures, foreign-control risks, human oversight, summaries of recorded processing steps, advance notice of material changes and model-version transition periods. Recorded-step summaries are not required to faithfully explain internal model reasoning. Government evaluation rights cover bias, truthfulness, safety and other factors. Failure to remediate after specific written notice can lead to termination for cause and reasonable decommissioning costs, capped at 25 percent of the affected task or delivery order and excluding specified replacement costs.

Related coverage

• Federal regulations and executive actions
• Government use and procurement
• NIST AI Risk Management Framework

References