From the State Government edition of October 5, 2026
Vermont oversight leaders' memo to ADS: the Anthropic evaluation agreement is silent on source-code retention and training, and may have skipped required contract approvals
Vermont General Assembly (Sen. Rob Plunkett, Rep. Laura Sibilia, Rep. Monique Priestley) to the Agency of Digital Services · State legislative IT oversight / AI cybersecurity tool contracting · Vermont statewide
- Publisher
- Vermont General Assembly (JITOC)
- Original publication
- 2026-09-24 (memo date; posted with Sep 30 JITOC materials)
- Source retrieved
- 2026-10-06
- Event date
- 2026-09-30
What happened
The Sep 24 memo from JITOC chair Sen. Plunkett and Reps. Sibilia and Priestley asks ADS for written answers on three issues: protection of state source code, the contracting process, and whether a NIST impact and risk assessment was done before the trial. It says Claude tools run in a customer environment but send code to an externally hosted model, so local operation does not by itself keep code inside state systems, and that the agreement shown to legislators has no provisions on retention, use, human access, model training or deletion of state source code. It cites Bulletin 3.5 requirements (sole-source justification approved by the CTO; Secretary of Administration and Attorney General review of zero-dollar contracts) and asks ten questions plus the executed agreement, approvals, pre-deployment assessment, and which hosting option (Anthropic, AWS, Microsoft Foundry or Google Cloud) Vermont used. The attached unsigned form provides $100,000 in credits for six months, limits use to defensive cybersecurity, provides services 'as is', labels Claude Security a beta service, and caps each party's liability at the greater of $1,000 or fees paid.
Why it matters
This is the document our Oct 4 edition listed as not located. It turns the Vermont dispute into a reusable checklist for any state taking free or credit-funded AI tools: data terms for code, sole-source and zero-dollar contract approvals, gift acceptance, and a documented pre-deployment risk assessment. Why now: ADS's written answers are due Oct 7.
Evidence and measured results
Opened the 7-page JITOC PDF over HTTPS (curl, UA SLED-research-bot/1.0) and extracted text with pdftotext; PDF metadata shows creation Sep 24, 2026. Confirmed the three concern areas, the Bulletin 3.5 citations, the ten questions and document requests, the hosting options question, and the evaluation agreement terms (credits, period, approved use, as-is disclaimer, beta label, liability cap).
Limitations and uncertainty
Advocacy document from the legislators raising the questions; it states concerns and requests, not findings of violation. The attached agreement is an unsigned standard form, so executed terms may differ. ADS's written answers were not yet available. Vermont Bulletin 3.5 and statutes are state-specific.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-05; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Free or credit-funded AI tools reach state code without the data terms and approvals a paid contract would get.
- Stakeholders
- Agency of Digital Services secretary
- State CTO and CISO
- Secretary of Administration
- Attorney General's office
- Legislative IT oversight committee
- Discovery questions
- How are zero-dollar AI agreements reviewed?
- Do your AI agreements cover retention, training and deletion of state data?
- Was a risk assessment done before data was shared?
- Which hosting option does each AI tool use?
- Value hypothesis
- A short review of AI trial agreements can find missing data terms and approvals before an oversight body does.
- Potential engagement
- Review of one AI trial's agreement, approvals and pre-deployment assessment, with a gap list and remediation plan.
- Unsupported claims
- That Vermont violated procurement law
- That state code was retained or used for training
- That Anthropic's terms are worse than other vendors'
- That a contract review would have prevented the dispute
Pre-sales engineering
Role takeaway
- Fit
- Pre-trial data-path and terms review for AI code-analysis tools.
- Architecture and integration
- Identify hosting path (vendor or cloud marketplace), scope repositories, scan for secrets before submission, and log what code was sent.
- Prerequisites
- Executed agreement with state attachments
- Inventory of repositories in scope
- Secret scanning
- Pre-deployment risk assessment
- Constraints
- Agreement shown was unsigned
- Vermont-specific Bulletin 3.5 rules
- Beta service with as-is terms
- Hosting option not disclosed
- Security
- Treat source code as sensitive: remove secrets, minimize what is shared, prefer hosting with contractual no-retention and no-training terms, and log every submission.
- Proof of value
- Data path documented and matched to executed terms
- Secret scan run on all submitted code
- Assessment dated before first use
Delivery
Role takeaway
- Work
- Compile the executed agreement, approvals, risk assessment and hosting details and answer the committee's ten questions.
- Dependencies
- Contracts staff
- Attorney General review
- Vendor confirmation of retention and hosting
- Ownership
- Agency of Digital Services secretary with contracts and legal staff.
- Skills and adoption
- Agency leaders need to know that free trials still trigger contract rules; developers need secret-handling practices.
- Governance checkpoints
- Sole-source or zero-dollar approval
- Attorney General review
- Pre-deployment risk assessment
- Committee disclosure
- Acceptance criteria
- Proposed: complete written responses by the committee deadline
- Proposed: every AI trial has an executed agreement on file
- Proposed: risk assessment dated before first data sharing
- Risks
- Executed terms mirror the as-is vendor form
- No assessment exists
- Other credit-funded trials not on the inventory
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Hosting choice matters: the same tool can reach Anthropic-hosted or cloud-marketplace models, which changes where code is processed and which data terms apply. Agencies should record which path is used.
Governance
Who approves, reviews and stays accountable for outcomes?
Shows that 'free' AI trials can still trigger sole-source, zero-dollar contract, Attorney General review and gift-acceptance rules, and that NIST alignment claims are expected to be backed by an assessment completed before use.
Security and privacy
What data, permissions and controls need testing?
Source code can expose system design, security controls, vulnerabilities and embedded credentials. The memo flags that the standard agreement lacks retention, training and deletion terms for that code.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Limited direct workforce content; implies agency leaders need training on when trials trigger contract and reporting duties.
Procurement
What should contracts, pricing and exit terms secure?
Directly relevant: zero-dollar and credit-funded AI agreements may need the same sole-source justification, CTO approval, Attorney General review and standard state contract attachments as paid contracts.
Operating model
Which teams own the service once it runs?
Oversight committees are asking for the contract file and risk assessment as the record of authorization, so agencies need a single owner who can produce it on request.
Publication history
- 2026-10-05State Government · Issue 305 resources
Stable resource ID: vermont-jitoc-memo-anthropic-evaluation-agreement-20260924