SDLC Playbook Security: How We Protect Your Evidence, Your Source, and Your Tenant
What we hold, where it lives,
and who can see it.
SDLC Playbook is an evidence product, so this page is written the way we would want a vendor to write one for us: specific, current, and honest about what is not done yet. Last reviewed November 15, 2026.
What we hold, and what we don’t.
We hold
Work-item and pull-request metadata (titles, identifiers, states, authors, reviewers, timestamps, linked requirements), pipeline and release outcomes, the governance rules you declare, the findings and evidence items the product generates, the SHA-256 hash and canonical record of each evidence item, chat approvals captured from Slack or Teams, and the documents the product authors for you.
We read but do not retain
Source code and diffs are read at analysis time for the agents you enable (Code Sentinel gates, the component map, documentation authoring). The product stores the findings and the component map it derives, not a copy of your repository. Chat messages are not ingested; only the approval event you point the product at is captured.
We never hold
Your production data, your customers’ data, credentials for your systems beyond the scoped connector tokens you grant, or payment card data. Your tracker and your repository remain the systems of record; the product holds the governance record about them.
Where it runs.
Hosting
Microsoft Azure, United States regions only. Azure App Service for the API, Azure SQL for data, Azure Service Bus for background work, Azure Key Vault for secrets, Azure Storage for generated documents and export packages. We inherit Azure’s physical, network, and platform controls and their SOC 2, ISO 27001, and FedRAMP authorizations; those authorizations cover Azure, not SDLC Playbook, and we do not claim otherwise. Azure Government deployment is on the roadmap and not available today.
Tenant isolation
Every table that holds customer data is protected by database row-level security keyed to the tenant. Queries that do not carry a tenant context return nothing. Isolation is enforced in the database, not only in application code.
Encryption
TLS for every connection in and out. Data at rest is encrypted by Azure SQL Transparent Data Encryption and Azure Storage service-side encryption. Connector tokens are held in Key Vault and referenced from the database, never stored there in plain text.
Identity and access
Customer sign-in runs on Microsoft Entra External ID. Services authenticate to each other with Azure managed identities; there are no shared service passwords. Administrative access to production requires multi-factor authentication, is limited to the founder, and is recorded in Azure activity logs.
What the product can touch, and what the models see.
Connector scopes
GitHub, Azure DevOps, Jira, Slack, and Microsoft Teams are connected with scoped access; the exact access for each is listed under Access at a glance below. Enforcement agents read. Authoring agents write only the artifact types you enable (a drafted story, a test plan, a comment) and never change work-item status. Every agent runs under a named roster entry with an explicit tool-access policy; an agent cannot call a tool its policy does not grant, and a deny-all policy switches an agent off. You can revoke any connector from the platform side at any time.
AI models
The product uses large language models for authoring, documentation, and analysis. Models are selected per tenant from a validated registry; each model is evaluated against prompt-injection and benign-refusal test sets before it is allowed in production, and the results are recorded. Secrets and personal data are redacted from content before any model call, and the redaction is logged as evidence. Every model call is metered and visible to the tenant. Models are accessed through commercial API terms that exclude training on customer content. Enforcement verdicts and the Acceptance Verifier’s checks are deterministic and involve no model.
What leaves your tenant, and what never does.
Every outbound path the product uses, with what crosses it and when it is on. If a path is not on this list, the product does not use it. Checked against the code on November 5, 2026.
GitHub, Azure DevOps, Jira
Reads pull-request, commit, work-item, pipeline and release metadata, and CI artifacts such as SonarQube and Snyk result files. Writes status checks, pull-request comments, and the drafted artifacts you approve. Receives their webhooks. On when you connect the platform, off when you revoke the connector. Source code is read at analysis time and never copied out of the platform.
Model providers
The text an enabled agent is working on, after secrets and personal data are redacted and the redaction is logged as evidence. Goes to Anthropic’s API under commercial terms that exclude training on customer content, or to your own Azure OpenAI deployment when you bring your key (today, the Coach path). Those are the only two provider transports: there is no Azure AI Foundry path, and the router does not fall back across providers. Every call is metered and the provider that answered is visible to you. Enforcement verdicts and the Acceptance Verifier make no model call.
Slack and Microsoft Teams
Out: notifications and approval requests to the channels you choose. In: the approval a person gives in reply, captured as evidence. Chat history is not read. On when you connect the workspace. Commercial cloud only.
Sign-in
Microsoft Entra External ID handles customer sign-in, so your users’ identities and sign-in events pass through Microsoft’s identity service. No other identity provider is used.
Platform telemetry
Performance and error telemetry goes to Azure Application Insights inside SDLC Playbook’s own Azure subscription, United States regions. Model-call usage records carry counts, model, latency and outcome, never content. Application logs are request metadata (method, path, status, duration), with no request or response bodies, and are being reviewed to the same no-content standard.
Nothing else
The product sends no email today; if an email sender is ever added, this line will name the service and what it carries. No vulnerability-database lookups, no third-party error tracking, no analytics inside the application, no web ingestion. Credentials never leave Azure Key Vault, and no path on this list carries your production data or your customers’ data, because the product never holds either.
What each connection can do, and what we keep.
For a team deciding whether to grant access to its repositories. Checked against the code on November 15, 2026. Anything we could not confirm from the code is left out.
| Connection | Reads | Writes | Worth knowing |
|---|---|---|---|
| GitHub GitHub App |
Repositories and branches. Repository contents: CODEOWNERS, the file tree and dependency manifests. Pull requests, reviews, changed files and commits. CI runs and artifacts. Code scanning, secret scanning and Dependabot alerts. | Check runs carrying the gate results, and Coach comments on pull requests. | The product only reads repository contents; no code path writes to your files or branches. It does not read branch protection or other administration settings. Access uses short-lived installation tokens that are never stored. GitHub shows the app’s full permission list on the install screen before you approve it. |
| Azure DevOps Entra sign-in or access token |
Code, pull requests, work items and builds. | Pull-request statuses and Coach threads. Work items, only for drafts a person has approved. | Azure DevOps offers no granular scopes for Entra sign-in: it grants the product delegated access to whatever the connecting user can do. For least privilege, connect with a personal access token limited to Code (read and write), Work Items (read and write) and Build (read). |
| Jira OAuth 2.0 |
Issues and projects, and user names (read:jira-work, read:jira-user). | Creating and updating issues (write:jira-work), only for drafts a person has approved. | The connection also requests offline_access so it stays connected between sessions. |
| Slack Bot |
Member lookup to verify a linked identity (users:read). | Posting notifications and approval requests (chat:write). | No history scopes are requested, so the product cannot read message history. Approvals arrive as the button press itself. The bot also requests channels:read, which the product does not currently use. |
| Microsoft Teams Bot |
Only activities addressed to the bot. | Posting notifications and approval requests. | No Microsoft Graph permissions are requested. The bot cannot read channel or chat history. |
Stored from a pull request
Title, state, source and target branch, head commit SHA, author login and email, who merged it, timestamps, the AI-involvement class with the ids of the rules that matched, and gate results with reviewer logins. Not stored: the pull-request description, diffs, commit messages, or file contents.
Tokens
Connector tokens live only in Azure Key Vault, encrypted under a key that is separate for each tenant. The database holds a reference, never the token. GitHub installation tokens are not stored at all.
Evaluating before you enforce
Observe mode records every verdict but posts each check as neutral, so nothing blocks a merge while you evaluate. It still needs write access to post those checks and comments. A read-only installation is not available today.
Proof that the proof wasn’t edited.
Every evidence item is serialized to canonical JSON and hashed with SHA-256 at the moment of capture. The hash is stored with the item and can be recomputed on demand through a verify endpoint, so an auditor, or you, can confirm an item has not changed since it was recorded. Audit export packages ship with a manifest that covers every file in the package, so a package can be checked for completeness as well as integrity. We do not provide a way to edit or delete an evidence item in place; corrections are recorded as new items that reference the original.
Backups, retention, incidents.
Backups
Azure SQL automated backups with point-in-time restore on Azure’s default short-term window; no long-term backup retention is configured. Azure Storage redundancy for generated documents and export packages, with a 7-day soft-delete. Deleted Key Vault secrets stay recoverable for up to 90 days in production.
Retention and deletion
Evidence and model-call usage records are kept for the life of your tenant; nothing expires them automatically today, and the per-plan retention periods are not yet enforced in code. Application logs are kept for 90 days. Disconnecting a connector stops new events and deletes its stored tokens from Key Vault; the records already collected remain. For GitHub, uninstall the app on GitHub to stop deliveries. When you leave, you can export everything first. Deletion is a manual operation today: we delete tenant data within 30 days of a written request and confirm in writing, after which backup copies age out under the windows listed under Backups.
Incidents and disclosure
We notify affected customers of a confirmed security incident involving their data within 72 hours of confirmation. Security researchers can report a vulnerability to info@sdlcplaybook.com; we acknowledge within two business days and do not pursue good-faith research.
What we have, what we’re doing, what we’re not.
Today. SDLC Playbook runs on itself: our own repository, pipeline, and releases are governed by the product, and the evidence on this page is the evidence we would show an auditor. We maintain a NIST SSDF (SP 800-218) practice posture generated by the product and will share it with any prospective customer under NDA. We have not completed a SOC 2 audit or an independent penetration test, and we do not hold ISO 27001, HIPAA, FedRAMP, or CMMC attestations. Any page that said otherwise would be wrong.
Planned. We intend to pursue SOC 2 Type I alongside our first paid customers, with Type II following the observation window, and an independent penetration test before Type I. The evidence for both will be collected by SDLC Playbook.
For your security review. We will complete your questionnaire (SIG Lite, CAIQ, or your own) and provide the architecture and data-flow diagrams on request. Write to info@sdlcplaybook.com.
Ask us anything on this page.
If a security reviewer needs something that isn’t here, we’d rather answer it than have you guess.