Stay connected: follow us on LinkedIn and explore more at
www.CherryHillAdvisory.com.
.svg)
Subscribe now to join the Risk Register community:
Source: DevAgentRadar (aggregated GitHub Copilot release notes)
Summary: DevAgentRadar’s aggregation of GitHub’s Sept 1, 2026 release notes highlights enterprise changes auditors can govern: Anthropic’s Claude Fable 5.1 is now available in Copilot Pro+, Max, Business, and Enterprise. Copilot code review adds an approval assessment and—if admins enable it—can submit approvals that count toward required approvals (public preview). Enterprises can set default Copilot models per org/team and manage deprecations with suggested alternatives. Budgets, model defaults, and a Customize tab (MCP servers, plugins, skills, canvases) are generally available across Copilot apps and IDEs.
How Internal Audit Can Deploy This: Treat these controls as levers to reduce IT risk in audit-built code and automation. Three ways to deploy now: 1) Control testing scripts: Set enterprise model defaults for audit repos so generated diffs and test logic are consistent across reviewers. Use code-review approval assessments to triage pull requests on control-testing utilities; require at least one human approver before merge. 2) Continuous monitoring and evidence: Enable Copilot budgets for audit teams; log spend and approval-assessment outputs as workpaper evidence of review for ICFR scripts and data utilities. 3) Technical risk governance: Use model-deprecation guidance to migrate audit tooling; document model choices in the risk register with rationale and fallback. Pilot steps this quarter: choose one low-risk repo with audit automation; set an enterprise/team default model; enable budget caps and alerts; restrict the Customize tab to approved MCP servers/plugins/skills; enable the approval-assessment preview on that repo. Run 3–5 controlled PRs (seeded with minor defects) and assess: precision of comments, false positives, time saved, and whether approvals were properly blocked without a human sign-off. Safeguards: prohibit secrets in prompts; keep agents read-only to production systems; restrict Customize tab to least-privilege connectors; require one or more human approvals in branch protection rules; archive approval-assessment output alongside PRs; align configurations to your AI governance policy and document admin changes.
✎ Try it yourself — Build recipe
Governed Copilot code‑review pilot for audit repos (Business/Enterprise) 1) Trigger/scope - Scope one non‑SOX, low‑risk repository containing audit scripts/utilities. Create a feature branch for test PRs. 2) Enterprise model defaults - In Copilot admin, set an enterprise/team default model for audit engineering. Map any deprecated models to the suggested alternatives and capture the decision in workpapers. 3) Budgets (GA) - Enable budgets for the audit team. Set a monthly cap and alert thresholds (e.g., 50/75/90%). Route alerts to the audit tools channel. 4) Customize tab governance (GA) - Approve only required MCP servers, plugins, skills, and canvases. Disable unapproved entries. Document the approved list and access rationale. 5) Repo protections + approval assessment (public preview) - Turn on Copilot code‑review approval assessment for the pilot repo. - In branch protection, require ≥2 approvals and explicitly require ≥1 human approver so Copilot approvals alone cannot merge. 6) Team overrides - If needed, set a team‑level default model for the audit group to enforce consistency across IDEs. 7) Test procedure - Submit 3–5 PRs with small seeded issues (naming, null checks). Record Copilot feedback, any Copilot approval, human reviews, and budget consumption. 8) Human approval gate - Code owners must sign off on merges. Reject merges that lack a human approval even if Copilot approved. 9) Go/no‑go - Go if: zero auto‑merges without a human; budgets/alerts fire; assessments are materially accurate. No‑go if any auto‑merge occurs or assessments materially miss seeded defects.
Source: Databricks documentation
Summary: Databricks (Beta) lets a Genie Agent analyze unstructured files in Unity Catalog volumes—PDFs, Word, slide decks, images—by attaching up to 10 volumes per agent. The capability runs only in Agent mode and uses multi‑step reasoning to retrieve and analyze relevant files. If a volume has content search, the agent queries the index rather than parsing full files, which the docs say avoids Foundation Model Serving retrieval costs. Workspaces must enable the “Analyze Files in Volumes with Genie Agents” preview. The docs show combined queries across documents and structured tables in one conversation.
How Internal Audit Can Deploy This: This is a pragmatic way to connect governed evidence to analysis without exporting files. Deploy it where audit needs both documents and tables in a single review. Use cases: 1) Evidence gathering for ICFR: Ask the agent to pull control descriptions from policy PDFs and cross‑reference execution logs or GL tables, returning citations to specific pages. 2) Full‑population analytics + document corroboration: Compare vendor contract terms stored in volumes to AP tables to flag payments that violate discount or renewal clauses; return the contract page as evidence. 3) Risk assessment: Scan quarterly board decks and risk memos in volumes, surface emerging risks, and tie them to entries in the risk register for follow‑up. Pilot this quarter: enable the workspace preview, create a dedicated audit agent, attach 1–3 read‑only volumes (policies, contracts, board decks), and ensure content search is enabled on those volumes to limit retrieval costs. Seed 10–20 test questions that require both documents and a known table. Require human review of extractions before using results in workpapers. Safeguards: restrict to least‑privilege volumes; keep the agent read‑only; log all prompts/responses; include citations (file path, page) in outputs; prohibit exporting files outside Databricks; align the setup to your AI governance policy and records retention rules.
✎ Try it yourself — Agent skill
Agent skill: Internal Audit Evidence & Corroboration (Databricks Genie) Role - You are a read‑only internal audit evidence agent operating in Databricks. You analyze documents in Unity Catalog volumes and corroborate with structured tables. Tools & boundaries - Access: Up to 10 attached Unity Catalog volumes (read‑only) and specified workspace tables/views. - If a volume has content search enabled, use the index first; do not re‑parse files unnecessarily. - No external web access. No data export. Respect table/volume permissions. Procedure 1) Parse the user question into: a) document targets, b) table fields/joins, c) required outputs. 2) Locate relevant documents via content search (prefer index). Retrieve only the top N snippets per file (N≤3) with page numbers. 3) Extract facts with citations (file path, page/slide, snippet). Note gaps/ambiguity. 4) Query referenced tables to corroborate facts (e.g., vendor_id, payment_date, amount). Join on obvious keys; summarize exceptions. 5) Produce a result set and an evidence log. Hard constraints - Read‑only. Never alter files or tables. Never invent facts. Always cite file path and page/slide. Stop and ask for clarification if joins or identifiers are ambiguous. Human escalation triggers - Missing citations, more than 10 documents implicated, or conflicting sources. Output format - Section A: Findings table (key, condition, result, table rows impacted). - Section B: Evidence table (file_path, page/slide, snippet, relevance note). - Section C: Assumptions and unresolved gaps.
Source: Google Cloud Blog
Summary: Google Cloud announced new billing options and cost controls for agent workloads in Gemini Enterprise and related developer tools: pay‑as‑you‑go alongside per‑seat subscriptions, pooled project quotas, and project‑level spend caps that can pause agent calls at budget thresholds. Also available are Flexible Savings Plans with one‑ and three‑year commitments and stated discounts, an AI Cost Summary/FinOps explainability agent for analyzing Gemini API and Vertex AI spend, automated budget alerts, and deferred‑execution pricing for off‑peak workloads.
How Internal Audit Can Deploy This: These controls close a common gap: agents that run long or variably and overrun budgets. Make them part of IT risk governance and, where relevant, SOX compliance around spend authorization. Use cases: 1) Control design/testing: Implement project‑level spend caps that pause agent calls at thresholds; test that caps trigger and halt execution. 2) Continuous monitoring: Use automated alerts and the AI Cost Summary/FinOps explainability agent to segment spend by project and driver, and reconcile to approved budgets. 3) Planning: Evaluate Flexible Savings Plans for stable agent workloads; keep variable or exploratory use on pay‑as‑you‑go. Pilot now: create a non‑production project with pooled quotas; set a monthly cap and 50/80/95% alerts; schedule non‑urgent agent jobs during off‑peak windows to leverage deferred‑execution pricing; run the explainability agent weekly to produce a cost variance report for risk oversight. Safeguards: restrict billing‑role access; require management approval before committing to Savings Plans; ensure caps and alerts are owned by a function separate from developers; document thresholds, pause behavior, and variance explanations in workpapers; align settings with your AI governance policy and budgets in the risk register.
✎ Try it yourself — Build recipe
Build recipe: Hard guardrails for AI agent spend (Google Cloud) 1) Trigger/scope - New or existing agent project in Gemini Enterprise/related tools with variable usage. Use a non‑production project for the pilot. 2) Billing mode - Select pay‑as‑you‑go for the pilot to measure baseline usage; retain per‑seat where already mandated by policy. 3) Pooled quotas - Enable pooled project quotas so teams share capacity without per‑user overages. Document expected daily/weekly usage bands. 4) Spend caps with auto‑pause - Set a monthly project‑level spend cap and alerts at 50/80/95%. Configure caps to pause agent calls when the threshold is reached. 5) Alerts & routing - Route budget alerts to a monitored channel (FinOps + Internal Audit). Define on‑call ownership to investigate overruns within 24 hours. 6) Savings plan guardrail - For steady workloads, evaluate one‑ or three‑year Flexible Savings Plans. Require a formal approval memo and budget owner sign‑off before purchase. 7) Deferred‑execution windows - Schedule non‑urgent agent jobs during off‑peak times to use deferred‑execution pricing. Label these jobs for reporting. 8) Cost explainability - Run the AI Cost Summary/FinOps explainability agent weekly to produce a report of spend by project and driver (e.g., model/API). Archive reports. 9) Test procedure - Generate controlled agent calls to cross the first alert threshold; verify alerts and that caps pause further calls at the cap. 10) Go/no‑go - Go if alerts fire, caps reliably pause calls, and explainability reports reconcile to billing. No‑go if any cap fails to pause workloads or reports cannot be reconciled.
Disclaimer: This content is provided for general informational purposes only and does not constitute legal, accounting, tax, investment, or other professional advice. Portions were generated using AI tools from public web sources and may contain errors or omissions — verify important details against the primary sources linked above before relying on them. Any example prompts, scripts, templates, or other artifacts are provided “as is” without warranty of any kind, express or implied; test them on non-production data and apply your own professional judgement before use. Cherry Hill Advisory disclaims all liability for any loss or damage arising from the use of, or reliance on, this content or any artifact it contains.
Subscribe now to join the Risk Register community:
