Now
China Turns Model Weights Into a Controlled Export Up to 0.5ptIMF estimate of AI’s possible yearly boost to world growth, over time Brazil Orders Meta’s AI to Treat Both Candidates Alike 33Dubai government entities in its agentic AI accelerator Abu Dhabi Puts Its Government Platform Up for Licence 1.85mcombined audience of a fake NATO-deaths video before Latvia’s vote Seven Days to Bid for India’s Sovereign GPUs 27 OctIsrael votes under its first deepfake-disclosure law The States Sue Where Congress Won’t 54risks listed in China’s AI Safety Governance Framework 3.0, up from 30 GSA’s AI Clause Goes Live as Washington Moves to Buy Agents €30bn+EU call for up to seven AI gigafactories Robocalls, Deepfakes and the Parody Defence 554deepfakes logged in Brazil’s first round; only 371 labelled Brasília Holds Its Server-Farm Tax Break Until the Run-Off 30+AI firms sent the first AI Act information requests 164AI-made ads in the US midterms, by one count 42signatories to Canada’s voluntary data-centre principles 174MWCape Town data centre now under appeal 83 dayswait reported under Medicare’s AI pilot, against a 72-hour standard 20home-grown foundation models backed by the IndiaAI Mission 2%of Turkish public investment budgets earmarked for AI

GSA’s AI Clause Goes Live as Washington Moves to Buy Agents

From 19 October, new GSA contracts for chatbots and agentic systems carry clause 552.239-7001: 72-hour incident notice, no training on government data and a record of each agent's steps. The Army and OPM show why the timing matters.

The Machine State desk
The pale stone General Services Administration headquarters building in Washington, DC, seen from a street corner under a clear blue sky
The General Services Administration headquarters building, Washington, DC, June 2024. Photo: ajay_suresh, CC BY 2.0, via Wikimedia Commons. (file page)

From today, Washington’s central buying agency writes its rules for purchased AI into new contracts. The General Services Administration’s class-deviation memo RGO-2026-01, from Senior Procurement Executive Jeffrey A. Koses, lists 19 October 2026 as the effective date of GSAR Part 552, which holds clause 552.239-7001, “Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems”. New GSA solicitations and contracts for a chatbot, generative assistant, agentic system or LLM-enabled productivity tool that handles government data now carry it; for awarded contracts, adoption is the contracting officer’s call. AI & GPP set out its scope on 12 October. The context has shifted: federal buyers are moving from chatbots that answer to agents that act.

The clause turns AI risk into contract deadlines

Most of what bites is a clock, set out below from the clause text.

TriggerVendor’s deadlineParagraph
Incident affecting government data (as defined in the Federal Information Security Modernization Act of 2014)Within 72 hours of discovery; daily updates; logs kept at least 90 days(f)(5)
Material violation of the clause72 hours from actual knowledge(d)(2)
Change that raises bias, weakens guardrails or degrades truthfulness7 calendar days(i)(3)
New or replaced model or provider30 days’ notice, unless concurrent access is given(i)(1)
Model retiredReasonable efforts to give 30 days’ (major) or 15 days’ (minor) access to the successor(i)(2)
Which models touch government dataDisclosed within 120 days of starting work, by default(f)(1)
Contract endsDelete data, fine-tuned weights, embeddings, indexes and caches(e)(4)(vii)
Termination for causeDecommissioning liability capped at 25 per cent of the order(j)(3)

The data-use limits are as blunt. Vendors may use government data to do the work, provide technical support, run de-identified security telemetry or as the contracting officer authorises in writing. Training any model on it, marketing or sales uses and selling it are prohibited. The buying government may also benchmark the deployed model at any time, including for “unsolicited ideological content”, and suspend its use.

Quinn Anex-Ries of the Center for Democracy & Technology called the final terms “a significant improvement from prior draft versions”, Nextgov/FCW reported; Jessica Tillipman of George Washington University greeted the removal of most “unbiased AI principles” language with “Hallelujah”, according to FedScoop.

The agent provisions matter more than the chatbot ones

Where a system relies on “reasoning, retrieval, or agentic steps”, the vendor must expose a summary of what it recorded, from routing choices and tool calls to retrieval sources and any human reviewer’s access, through the interface, an API and an audit trail. The clause asks only for “a summary of recorded steps, not a faithful explanation of the model’s internal reasoning.”

Who is already buying agents, and under whose terms

At the Association of the United States Army (AUSA) annual conference on 12 October, the Army’s principal cyber advisor, Brandon Pugh, said the “take action” element of Project Griffin, an autonomous response “within prescribed parameters of a human operator”, would be awarded that Friday, with contracts going to two companies, Breaking Defense reported. According to Breaking Defense, the August commercial solutions opening drew 108 submissions, cut to ten finalists. Lt Gen Christopher Eubank, head of Army Cyber Command, put the open question plainly: “Are we as humans ready to let the agent actually take action?” By 19 October the desk had found no announcement naming the winners.

Some agents are already inside. Task Force Lexington, formed in April, has moved agents out of a sandbox and into the Army’s part of the Department of Defense Information Network, with humans checking their work, DefenseScoop reported. Eubank said the command had been deliberate about its guardrails, including asking “how do we incrementally reduce the guardrails over time?”

These are Army buys. RGO-2026-01 applies to GSA’s own contracting, so, on this desk’s reading, Griffin’s terms are whatever the Army writes, not 552.239-7001.

On the civilian side, the Office of Personnel Management opened FWD Chat in public beta on 7 October. It turns plain-English questions into SQL queries on federal workforce records and shows the query behind each answer, FedScoop reported; OPM plans to link it to America.gov, the chatbot AI & GPP examined on 6 October. The FWD Chat page warns that “Responses may contain errors or vary between similar questions.” Publishing the SQL is a voluntary version of the traceability GSA now demands.

What decision-makers should do and watch

Contracting officers with open AI solicitations should decide now whether to add the clause. Vendors should test whether incident playbooks meet 72 hours and daily updates; for FedRAMP-authorised systems, reports filed under FedRAMP or to the Cybersecurity and Infrastructure Security Agency (CISA) count if they carry substantially the same information and reach the contracting officer at the same time. AI & GPP’s 12 October cover story compares other governments’ clocks.

Watch three things: the Griffin winners and their terms; whether defence buyers copy GSA’s step-trace and data-use limits; and 30 November, when changes to Part 538, which governs the Federal Supply Schedules, take effect. Formal rulemaking, with notice and comment, is still to come.

Sources

  1. RGO-2026-01, GSAR Class Deviations in Response to Executive Order 14275 (incl. clause 552.239-7001) — General Services Administration, memo dated 15 January 2026; Part 552 signed 16 July 2026, effective 19 October 2026; clause dated September 2026 (PDF)
  2. GSA memo to set AI-specific acquisition rules — Nextgov/FCW, 1 October 2026
  3. ‘Hallelujah’: Stakeholders react with praise and concern to GSA AI acquisitions clause — FedScoop, 6 October 2026
  4. Army to take a (cautious) step toward employing AI agents to defend networks: Official — Breaking Defense, 13 October 2026
  5. ARCYBER expanding use of AI agents for network operations — DefenseScoop, 12 October 2026
  6. OPM launches beta AI chatbot for federal workforce data — FedScoop, 9 October 2026
  7. FWD Chat – Federal Workforce Data — US Office of Personnel Management, undated (beta; data as of August 2026)

AI & GPP reports on how artificial intelligence and automation are changing the way governments decide, regulate and campaign. Corrections and tips: contact the editors.

Get the stories that matter to decision-makers, weekly.

Discussion

No comments yet. Start the conversation.

Discussion is open to members.

Join free Sign in

Read next

Abu Dhabi Puts Its Government Platform Up for Licence

Abu Dhabi has created TAMM Global to license its AI-driven public-services platform to other governments, with an American city tipped as the first customer. The pitch rests on self-reported results and on conditions at home that no licence can transfer.

Seven Days to Bid for India’s Sovereign GPUs

C-DAC opens a seven-day window on 22 October to empanel providers for up to 10,000 sovereign AI compute units. The entry bar is high, and New Delhi's promised data-centre policy still exists only as a ministerial remark.

No Register, No Rollout: Peru’s Health Ministry Sets Its AI Terms

A ministerial resolution bars Peru's health ministry from buying or deploying any AI system it has not registered, risk-rated and authorised. Elsewhere in Latin America, a Mexican electoral institute and Ecuador's legislature have just declined to write AI rules of their own.