Claude Security and Governance consulting

Security teams are usually asked a vague question: is Claude safe? The useful questions are narrower. Anthropic's Commercial Terms of Service (effective June 17, 2025) say Anthropic may not train models on Customer Content from the services, and that the customer owns its outputs. That part is settled for Team, Enterprise and API use. What is not settled by default is everything your company controls. On Claude Enterprise, chats and projects are kept indefinitely unless an owner sets a custom retention period. The audit log export reaches back 180 days. Connectors reach whatever the credential behind them can reach. Those defaults are reasonable for a pilot and wrong for most records schedules, so the review has to end in configuration, not a memo.

1. 01

   What the review covers

   We work through seven areas and document a decision for each: data use and retention, identity and provisioning, audit and monitoring, network access, connectors and MCP servers, the deployment route (Claude apps, the Claude API, Amazon Bedrock, Google Cloud or Microsoft Foundry) and regulated data such as PHI. Plan choice is part of the security answer. According to claude.com/pricing as of October 26, 2026, the Team plan includes SSO, domain verification and admin controls for connectors, but SCIM, audit logs, the Compliance API, custom data retention, role-based access, inference hooks, IP allowlisting and the HIPAA-ready offering are Enterprise features. If your requirements include any of those, the license decision is made before procurement, not after.

2. 02

   Where Claude governance usually breaks

   The failures we look for are specific and documented by Anthropic itself. A custom connector set up with a fixed API key gives every member that key's access, so a broad key turns a read-only use case into a company-wide data path. Shortening a retention period schedules everything outside the new window for permanent deletion when you save, with no recovery. Resetting SCIM deletes every Claude account except the Primary Owner. Turning on IdP group mappings can demote an admin who is not in a mapped admin group. Enabling HIPAA is one-way from organization settings. None of these are exotic. They happen when settings are changed by one person without a written change plan, which is why the register comes first.

3. 03

   Why Abstrakt Solutions

   We have been a Salesforce partner since 2017, and most of that work is access design: profiles, permission sets, sharing rules, field-level security and integration users that see only what they should. Claude governance is the same discipline applied to a new surface, and it matters most when Claude connects to the CRM through Claudeforce or an MCP server. Our team holds 12 Claude certifications, and Abstrakt Solutions is a Certified Services Partner in the Claude Partner Network. This engagement covers company-wide controls for Claude. If you need approval steps inside a regulated workflow, such as a disclosure review or a claims decision, that is our separate Claude compliance automation service.

The engagements we run most often on Claude Security and Governance, from first implementation through optimization.

1. 01

   Inventory current AI use and data classes

   We list who uses which AI tools today, including personal accounts on company email, and classify the data each team wants to bring to Claude: public, internal, confidential, regulated. We pull your records schedule, current DLP rules and vendor-risk questionnaire. Owners: IT security and legal, with one lead from each pilot department. The output is a one-page data map that drives every later setting, including which teams may connect which systems.

2. 02

   Choose plan and deployment route

   We compare what each requirement needs against the current plan matrix and the cloud routes. Examples: SCIM, audit logs, custom retention and the HIPAA-ready offering require Enterprise; a team building internal apps on its AWS account may be better served through Bedrock with retention locked by a service control policy. We confirm open items with your Anthropic or cloud account team in writing before signature, since contract terms override any help-center summary.

3. 03

   Configure identity, roles and network

   We verify your domains, turn on restrict organization creation, connect SSO and choose JIT or SCIM. Before saving provisioning, we confirm every intended user is assigned to the Anthropic app in your IdP, because unassigned users are deprovisioned on save. We build custom roles per group, set model and capability access, and submit IP ranges for allowlisting, including VPN exit points. Owners: your identity admin and our consultant, working from a change plan.

4. 04

   Set retention, audit and connector controls

   We set chat, project and artifact retention to match the records schedule, export a baseline audit log, and stand up the Compliance API feed to your SIEM or archive with read and delete scopes held by different people. Each connector gets a named business owner, a scoped credential and per-tool permissions by role. If you use inference hooks, we start in shadow mode and review verdicts with security before enforcing them.

5. 05

   Publish policy, train and rehearse an incident

   We publish the acceptable-use policy with a settings register that shows which control enforces each rule, then brief managers on what changes for their teams. We run a tabletop exercise: an employee pastes customer data into the wrong place, or a connector starts returning unexpected content. The exercise tests who pulls the audit trail, who disables the connector and who tells legal. We schedule a quarterly review, since Anthropic ships admin changes often.

1. 01

   Treating 'no training' as the whole data answer

   The training clause is settled, but retention, deletion, audit and connector reach are not, and Enterprise keeps data indefinitely until someone sets a period.

2. 02

   Saving provisioning before checking IdP assignments

   Users not assigned to the Anthropic app when JIT or SCIM settings are saved get deprovisioned, and admins outside mapped groups can lose rights. Test against a roster export first.

3. 03

   Adding connectors with one broad shared key

   A fixed request header gives every member the same access. Use per-user authentication or a narrowly scoped service credential, and keep write tools off until approved.

4. 04

   Assuming the BAA covers every feature

   Connectors, skills, Claude in Chrome and cloud sessions sit outside BAA coverage. Without a role that blocks them, staff can send PHI through a path the agreement does not cover.

Other Claude platforms we implement

• Salesforce sharing model explained: who sees which records and why
• Salesforce backup and recovery: what to protect and how to restore
• Salesforce single sign-on: SSO, MFA and user provisioning
• Experience Cloud security: guest users, sharing and private data
• Does Claude train on your data? Training, retention, zero data retention and where data lives
• Is MCP safe for business use? The real risks and the controls that work
• Salesforce for HR teams: employee service, cases and onboarding
• Salesforce data privacy compliance: CCPA, GDPR and design choices