From the State Government edition of October 8, 2026
Vermont ADS tells JITOC it will apply Bulletin 3.5 waivers to future AI trials, is not using Claude Security or Mythos, and has used about 10% of a $100,000 Anthropic credit
Vermont Agency of Digital Services (Secretary/CIO Denise Reilly-Hughes; Chief Data and AI Officer Josiah Raiche) · State IT cybersecurity code review and AI evaluation oversight · Vermont statewide (executive-branch ADS)
- Publisher
- Vermont General Assembly / Agency of Digital Services
- Original publication
- 2026-10-07
- Source retrieved
- 2026-10-09
- Event date
- 2026-10-07
What happened
On Oct 7, 2026, ADS posted formal answers to the Joint Information Technology Oversight Committee on its June 2026 Anthropic SLTT evaluation. Secretary/CIO Denise Reilly-Hughes says ADS entered a six-month evaluation to examine state system code for security and operational weaknesses; the evaluation is limited to system code and no personal information or Vermonters’ data has been reviewed. ADS did not bid, request a Bulletin 3.5 waiver, or treat the agreement as a procurement at signing, but after consulting the Secretary of Administration it will use the Bulletin 3.5 waiver process for trials and evaluations going forward. CDAO Josiah Raiche writes Vermont is not using Claude Security and does not use or have access to Claude Mythos; developers use the Claude Code plugin in VS Code with Sonnet, Opus, or Fable by application size, after a two-day in-person training. Sonnet/Opus prompts are not retained; Fable retains prompts and attached code up to 30 days for misuse review. All generated code goes through standard release review before production. ADS is using the Azure-hosted LLM, has used about 10% of the $100,000 credit allocation, and will add the use case to the state’s AI Inventory at the evaluation’s conclusion. Authorization was signed by the then-serving CISO and reviewed by ADS General Counsel and the CDAO.
Why it matters
Primary-source answers to legislative oversight on a free frontier-AI cybersecurity trial: procurement classification, model choice, data retention, human release review, and a prospective process change (Bulletin 3.5 waivers for future trials). Transferable as a template for how other states document SLTT-style evaluations under legislative scrutiny, with Vermont-specific Bulletin 3.5 and inventory rules as jurisdictional limits.
Evidence and measured results
Opened the Oct 7 ADS Response PDF from the JITOC documents page (Denise Reilly-Hughes cover letter, Raiche memorandum, Attachment A Q&A dated Oct 6) via curl and pdftotext. Confirmed listing on https://legislature.vermont.gov/committee/document/2026/345/Date/10-7-2026. Cross-checked Compass Vermont reporting (already archived) and the signed evaluation agreement posted the same day.
Limitations and uncertainty
Agency self-report under oversight pressure, not an independent audit. No count or severity of vulnerabilities found, no measured remediation outcomes, and no confirmation whether Fable was used on state code. Attachment A is dated Oct 6 while the cover letter is Oct 7. The agreement title/services include Claude Security even though ADS says it is not used—see companion contract resource.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-08; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Legislators are questioning free frontier-AI cybersecurity trials that skipped formal procurement paperwork.
- Stakeholders
- State CIO / ADS leadership
- JITOC or equivalent IT oversight committee
- Secretary of Administration / procurement
- CISO and CDAO
- Agency general counsel
- Discovery questions
- Was the AI trial processed under procurement bulletins or treated as a non-procurement evaluation?
- Which models and retention settings are authorized, and is Mythos or whole-repo scanning in scope?
- What human release-review steps sit between AI-suggested patches and production?
- When will the use case appear on the public AI inventory?
- Value hypothesis
- Documented, waiver-ready evaluation packages with model and retention controls can satisfy oversight while still testing AI code review.
- Potential engagement
- A time-boxed evaluation readiness review that maps a proposed SLTT trial to Bulletin-style waiver and inventory steps.
- Unsupported claims
- Measured reduction in vulnerabilities or mean-time-to-remediate from the Vermont trial
- That Claude Security or Mythos were used in Vermont
- That Bulletin 3.5 was followed for the original June agreement
Pre-sales engineering
Role takeaway
- Fit
- Fits agencies evaluating AI-assisted code review in test/dev with human SDLC gates.
- Architecture and integration
- VS Code plugin + Azure-hosted LLM; share limited folders; scale Sonnet/Opus/Fable by app size; keep Claude Security optional if stack-incompatible.
- Prerequisites
- Developer training
- Counsel review of vendor privacy/security terms
- Security channel for sensitive threat-model outputs
- Constraints
- No personal data in scope
- Local folder access model
- Fable 30-day retention on large apps
- Vermont Bulletin 3.5 and inventory rules are jurisdiction-specific
- Security
- Prefer zero-retention models where workable; treat threat-model docs as sensitive; confirm no Vermonter PII leaves the evaluation boundary.
- Proof of value
- Sample of AI-suggested findings with human disposition recorded
- Confirmation of model tier and retention setting per repo
- Release ticket showing human review before production
Delivery
Role takeaway
- Work
- Run a governed AI code-review evaluation with training, waiver/procurement path, and inventory close-out.
- Dependencies
- Procurement/admin waiver process
- Security Division review capacity
- Updated SDLC for higher patch volume
- Ownership
- ADS (CIO/CDAO/CISO), with legislative oversight visibility.
- Skills and adoption
- Two-day developer orientation plus Pathfinder ethics training for non-technical sponsors.
- Governance checkpoints
- Pre-trial counsel and leadership sign-off
- Model and retention authorization recorded
- AI Inventory entry at trial end
- Acceptance criteria
- Proposed: every production patch from the trial has a named human reviewer
- Proposed: waiver or equivalent procurement documentation filed for new trials
- Proposed: inventory entry lists vendor, free credit, and data-sharing status
- Risks
- Oversight finds paperwork gaps mid-trial
- Contract scope includes tools the agency says it is not using
- Retention on higher-tier models expands code exposure
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?
Claude Code runs as a VS Code plugin against folders on the developer’s machine; samples of code and file structure are shared with Anthropic APIs for orientation, threat modeling, and vulnerability suggestions. ADS uses Azure-hosted LLM access. Model tier scales with codebase size (Sonnet/Opus/Fable). Claude Security is not used because it requires whole-codebase access outside the existing stack.
Governance
Who approves, reviews and stays accountable for outcomes?
ADS reframes future AI trials as Bulletin 3.5 waiver matters after legislative challenge, while defending the original non-procurement characterization. Ongoing oversight is framed as use-case guardrails and AI Inventory/use-case review (NIST-informed) rather than only tool listing. Authorization was CISO signature plus counsel and CDAO review—not Secretary of Administration or AG Bulletin 3.5 approvals.
Security and privacy
What data, permissions and controls need testing?
ADS asserts no Vermonter personal data in the evaluation and that Mythos/Claude Security are not in use. Fable retention (up to 30 days of prompts and attached code) is accepted for large codebases; Sonnet/Opus have no retention. Threat-model outputs are treated as highly sensitive and shared with ADS Security when significant issues appear.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Two-day in-person developer training and Pathfinder training for non-technical staff frame AI as human-operated power tools; ADS says it does not deploy autonomous agents. Staffing levels are not discussed.
Procurement
What should contracts, pricing and exit terms secure?
Free $100,000 token credit via Anthropic’s SLTT program was treated as a non-procurement evaluation; ADS now commits to Bulletin 3.5 waivers for similar trials. Attachment C and AA-14 were not used. Vendors and agencies should expect legislative demand for competitive-process documentation even on zero-dollar AI evaluations.
Operating model
Which teams own the service once it runs?
Code-review workflow: orientation → threat model → vulnerability scan → developer-chosen Claude Code patches → SDLC release review (being updated for higher volume). Security Division is looped in for significant findings. Inventory entry deferred until evaluation ends.
Publication history
- 2026-10-08State Government · Issue 332 resources
Stable resource ID: vermont-ads-jitoc-anthropic-sltt-response-20261007