Security & governance
the complete guide

1. Data retention and zero data retention

Default retention policy

Anthropic's default retention window for commercial customers is 30 days. API inputs and outputs are deleted automatically. Conversations through Claude Team and Enterprise stay until a user explicitly deletes them, and once deleted they're purged from the backend within 30 days.

Exceptions:

• Usage-policy violations: up to 2 years
• Safety classification scores tied to violations: up to 7 years
• Legal or contractual hold: custom periods

Zero data retention (ZDR)

With ZDR, customer data is dropped from memory after the API response is returned and nothing is persisted — except for what's needed to meet law-enforcement or abuse-prevention obligations.

Scope	Covered / not covered
ZDR covers	Claude Messages API, Token Counting API, Claude Code (via API key and via Claude Enterprise)
ZDR does not cover	Console, Workbench (web UI), Claude Consumer (Free / Pro / Max), Claude Teams, Claude Enterprise UI (except Claude Code), third-party integrations such as Bedrock and Vertex AI

Enablement: contact the Anthropic sales team. They verify eligibility and turn ZDR on per organization. Switching it on automatically disables certain features that need conversation history to work.

The commercial-terms guarantee

Anthropic's commercial terms state in writing that customer content is not used to train models. Inputs and outputs from commercial customers never end up in the permanent model-improvement pipeline.

2. Identity and access (SSO, SCIM, MFA, RBAC)

Single sign-on (SSO)

Claude Enterprise supports SAML 2.0 and OIDC. Supported IdPs:

• Microsoft Entra ID (Azure AD)
• Google Workspace
• Okta
• Ping Identity
• Any standards-compliant SAML or OIDC IdP

Enforcement options:

• Require SSO for Console — admins must sign in via SSO
• Require SSO for Claude — end users must sign in via SSO
• MFA policy is inherited from the IdP automatically

SCIM user provisioning

SCIM directory sync is Enterprise-only. When you add or remove a user in the IdP, Claude follows within seconds. Remove a user in the IdP and their access dies across web, desktop, and CLI at the same time. That matters on the day someone leaves.

Role-based access control (RBAC)

Role	Permissions
Admin	All configuration and user management
Member	Standard Claude usage
Workspace-specific roles	Per-project delegation

Domain claim

Claim your corporate email domain and any signup attempt from that domain gets routed into the Enterprise workspace. It's the cleanest way to kill shadow IT on personal accounts.

3. Audit and compliance (SOC 2, ISO, HIPAA, GDPR, PIPA)

Certifications

Certification	Scope	Notes
SOC 2 Type I & II	Anthropic corp	12-month observation window, five trust criteria audited by an independent party
ISO 27001:2022	Information security management	API infrastructure, customer data processing, internal access control
ISO/IEC 42001:2023	AI management system	AI security and quality management
HIPAA-ready	API layer (BAA required)	For organizations handling protected health information

Enterprise customers can pull SOC 2 Type II reports and ISO certificates from the Anthropic Trust Portal under NDA.

GDPR and DPA

Anthropic provides a Data Processing Addendum to customers subject to GDPR. Standard Contractual Clauses are included. Accepting the commercial terms pulls in the DPA automatically.

HIPAA (healthcare)

Once a Business Associate Agreement is signed, the Claude API exposes a HIPAA-ready surface. Customers handling PHI get a separately provisioned HIPAA-enabled organization.

Note: Claude Code is not covered by HIPAA. API only.

Japan (APPI)

Anthropic processes personal information in line with Japan's Act on the Protection of Personal Information. The Japan entity is Anthropic Japan GK. Personal information is held on secure servers with encryption and access restrictions in place.

4. Japan data residency

Direct API (Anthropic-hosted)

Anthropic's direct Claude API (api.anthropic.com) runs in US and Global regions only. There's no Tokyo region. If you've got a legal or contractual requirement for in-country processing, the direct API won't meet it.

AWS Bedrock ap-northeast-1 (recommended)

Verified 2026-04-20: Claude Opus 4.7 is generally available in AWS Bedrock Tokyo (ap-northeast-1) for production use.

Regional behaviour:

• ap-northeast-1 — Tokyo data centre
• Cross-region inference profiles keep traffic within US, EU, Japan, or Australia
• Quota increases have to be requested from the region you need them in

Available models:

• Claude Opus 4.7 (latest)
• Claude Sonnet 4.5 / 4.6
• Claude Haiku 4.5

Google Cloud Vertex AI (Tokyo)

Vertex AI supports both multi-region endpoints (dynamic routing across regions) and regional endpoints pinned to a single region. Pin it to Tokyo and processing stays in Japan.

Note: when you integrate directly with Vertex, Google Cloud's DPA applies, not Anthropic's.

5. API key management — the essentials

We have a whole page on API keys. For the full walkthrough, see the API key hygiene guide. Here are just the governance-level bits.

Secret managers

Tool	Best fit	MIXI recommendation
AWS Secrets Manager	Dynamic fetch in Lambda, EC2, ECS	★★★★★
1Password	Team sharing plus audit logs	★★★★★
Doppler	CI/CD and env-var sync	★★★★☆
HashiCorp Vault	On-prem and hybrid	★★★★☆

Don't:

• Hard-code keys or commit them
• Paste them into Slack, email, or shared docs
• Keep long-lived keys with no rotation
• Share admin keys across people

Rotation schedule

Rotate every 90 days, or immediately on detected leak.

Staged rotation, five steps:

1. Generate the new key, store in Secrets Manager
2. Update app env vars to the new key (canary first)
3. Run old and new in parallel for 7 days
4. Disable the old key
5. Confirm retirement at the monthly review

Scope minimisation

One key, one purpose.

• PROD_API_KEY — production only, read/write
• STAGING_API_KEY — staging, read-only (optional)
• CI_CD_KEY — GitHub Actions or GitLab Runner, scoped to the project
• LOCAL_DEV_KEY — per-developer, 30-day expiry on dev machines

Audit logs

The Claude Enterprise dashboard records every key operation. Check monthly:

• Key creation and deletion history
• Anomalous call patterns — geography, rate, error percentage
• Unused keys flagged for retirement

6. Incident response — the 15-minute playbook

What counts as an incident

Any of these triggers a 15-minute response:

• API key accidentally committed to a repo
• GitHub public repo exposure caught by an automated scanner
• Cost spike — 5x or more week-over-week
• API calls from an unexpected geography
• Suspected prompt injection (Anthropic abuse report)
• Unauthorized tool execution or file access by Claude Code

The 15-minute playbook

Time	Owner	Action	Deliverable
T+0	Detector	Incident report in Slack #security-alerts	Short written report
T+3	Security lead	Escalate to Anthropic support (support@anthropic.com)	Support ticket ID
T+5	Dev lead	Revoke or disable the suspect key from Console immediately	Revocation confirmed
T+7	Auditor	Pull last 24h of API logs, check for anomalies (who called what)	Log analysis report
T+10	Security lead	Update all teams; decide on exec notification	Risk assessment
T+15	Named owner	Generate and deploy new keys; lock in the rotation schedule	Recovery plan

Anthropic contacts

• Support: support.anthropic.com
• Abuse report: api-abuse@anthropic.com
• Security issues: HackerOne anthropic-vdp

7. Prompt injection countermeasures

Known attack vectors

Per Anthropic's production research, roughly 90% of Agent incidents originate from tool-output injection.

Vector	Share	Example
Direct prompt	10%	A user types malicious instructions straight into Claude
Tool-output injection	50%+	Instructions embedded in API responses or Web fetch results
MCP server poisoning	20%+	An untrusted MCP server returns malicious payloads
Config and env poisoning	10%+	ANTHROPIC_BASE_URL overridden to an attacker endpoint; Hooks or MCP config tampered with

Claude Code defences

Permission-based architecture:

• Read-only by default — file writes and command execution require explicit permission
• /sandbox isolates bash runs (filesystem and network restrictions)
• Command blocklist — curl, wget, and other arbitrary-content fetchers are blocked by default

Cowork (remote collaboration) restrictions:

• Trust Verification prompt when cloning untrusted repos
• File watch — no writes above the parent directory
• Network request approval — external API calls need pre-authorisation

MIXI-recommended three-layer defence

Layer	Controls
Layer 1 — Model	Use Claude Sonnet 4.5 or later (improved attack detection)
Layer 2 — System	Allowlist MCP servers from trusted providers only; run Claude Code inside a VM or devcontainer
Layer 3 — Operational	Log all API calls to CloudWatch; monthly anomaly-detection dashboard; code review for risk assessment

8. Model training implications

Default behaviour

Anthropic does not train models on commercial customer inputs or outputs — it's in the commercial terms in writing. Two exceptions:

• User Safety classifier results (abuse-detection scores) — retained and used
• Aggregated, de-identified patterns — usable for improvement research

Enterprise opt-out

With ZDR, the guarantees tighten further:

• Inputs and outputs are deleted from storage after the API response returns
• User Safety classification is also deleted unless a policy-violation flag fires

Consumer vs commercial

Surface	Training use	Where to check
Claude Free / Pro / Max	Possible — user can opt out in privacy settings	claude.ai/settings/privacy
Claude API (commercial)	No	Automatic; confirmed in ZDR contract
Claude Code Desktop	No	Confirmed via org-level ZDR

9. Network and transport (TLS, IP allowlisting, mTLS)

TLS encryption

All traffic to the Claude API runs over TLS. Encryption strength is industry standard.

• API endpoint — api.anthropic.com (HTTPS only)
• Remote Control sessions — TLS over a streaming connection (local session routes messages through the Anthropic API)

Egress IP restrictions

Claude Enterprise supports IP allowlisting.

• Admins set an allowed-IP list in Organization settings
• Authentication attempts from client IPs outside that list get blocked

Example: only the MIXI HQ LAN segment 203.0.113.0/24 can authenticate.

Corporate proxy and mTLS

Enterprise deployments can configure the following via environment variables:

• HTTP proxy for outbound traffic
• Custom Certificate Authority (CA) trust
• mTLS client certificates for mutual authentication

Watch out for

• WebDAV on Windows is Microsoft-deprecated — avoid
• Package manager access (npm, PyPI) can be scoped to allowed domains, but token exfiltration risk remains — minimise credentials

10. Enterprise procurement

DPA and BAA signing

Step	Duration	Owner
Initial sales contact	Same day	Anthropic account team
Scope definition	2–3 days	Legal + InfoSec + Sales
DPA / BAA draft delivered	1–2 days	Anthropic Legal
Internal legal review	5–7 days	MIXI Legal
Redlines and negotiation	7–14 days	Anthropic Legal + MIXI Legal
Signature and execution	1–2 days	Approver and signatory

Standard DPA includes Anthropic's Standard Contractual Clauses.

InfoSec questionnaire / RFI

Anthropic provides:

• SOC 2 Type II report (under NDA)
• ISO 27001 and ISO 42001 certificates
• HIPAA Implementation Guide (when HIPAA is in scope)
• Penetration test summary on request

Recommended process:

1. Submit your security questionnaire to Anthropic (any standard template works)
2. Anthropic compliance team typically responds within 3–5 business days
3. Take the response to the internal information security board for review

Typical review timelines

Contract type	Timeline
Lightweight (no ZDR)	2–3 weeks
Standard (ZDR + baseline compliance)	4–6 weeks
Complex (HIPAA + custom SLA)	8–12 weeks

Start the Anthropic sales conversation early and kick off legal in parallel.

11. Approval fatigue — the human-in-the-loop problem

Every executive who tries Claude Code surfaces the same worry within a week. "I'm going to press Approve so many times that eventually I'll stop reading what's in the diff." It's not paranoia. It happens. Human attention drops off sharply inside a 30-minute session. Approval dialog number forty isn't being read — it's being clicked.

And that habituation is where the security risk lives. One backdoor line dropped into a larger diff gets Accepted without a glance.

Five mitigations

Exec KPIs

Make approval quality visible as numbers and abstract worry turns into an operational problem you can actually fix.

KPI	Target	How to measure
Post-merge defect rate on Claude-authored PRs	<2%	Monthly pull of merged PRs cross-referenced with the bug tracker
Post-approval revert rate	<5%	Claude-generated commits reverted within 30 days
Red-team detection rate	80%+	Share of monthly-drill traps that got caught
Average approvals per session	Under 20	High numbers mean Plan mode isn't doing its job

12. Official sources

Everything here traces back to the Anthropic documentation and Trust Center materials below. Last verified 2026-04-22.

• Anthropic Trust Center
• Privacy Center
• Claude Help Center
• API Documentation
• Claude Code Docs
• Legal Docs
• Security Report — HackerOne
• Zero Data Retention — Claude Code Docs
• AWS Bedrock Claude Opus 4.7 in Japan (2026-04-20)
• Set up Single Sign-On (SSO) — Claude Help Center
• Restrict Access with IP Allowlisting

Based on information as of 2026-04-21. Anthropic ships fast — features, pricing, and compliance scope can change without notice.