Skip to main content

AI Provider Data Security

Every AI provider ToolboxAI supports — OpenAI, Anthropic (Claude), and Salesforce Agentforce — is independently SOC 2 audited, and each has its own data retention policy and security architecture on top of that. This page summarizes what each provider does with your data, and what ToolboxAI itself does to reduce what leaves your org in the first place.

OpenAI​

  • SOC 2: OpenAI undergoes an independent SOC 2 Type II examination covering Security, Availability, Confidentiality, and Privacy for both the API and ChatGPT Enterprise products. OpenAI also holds ISO/IEC 27001:2022 and ISO/IEC 27701:2019 certifications. Reports are available to customers via the OpenAI Trust Portal.
  • Data retention: API data is not used to train OpenAI's models by default, and API inputs/outputs are retained for 30 days for abuse and safety monitoring, then deleted. Qualifying enterprise customers can request Zero Data Retention (ZDR), where inputs and outputs are not stored at all after the response is returned.
  • Unique measures: encryption in transit and at rest, and data residency options for enterprise customers who need requests processed in specific regions.

Anthropic (Claude)​

  • SOC 2: Anthropic has completed a SOC 2 Type II audit covering Claude API infrastructure and internal controls (report available under NDA via Anthropic's Trust Portal).
  • Data retention: standard API usage is not used to train Anthropic's models. Eligible enterprise accounts can request a Zero Data Retention (ZDR) addendum, under which prompts and responses are processed in real time and not written to disk after the response is returned — except where retention is required to investigate a policy violation (up to 2 years in that specific case) or comply with law.
  • Unique measures: in-pipeline abuse/safety checks that don't require persisting conversation data, and encryption in transit and at rest.

Salesforce Agentforce​

  • SOC 2: Agentforce is a covered service in Salesforce's Einstein Platform SOC 2 and SOC 3 reports, and Salesforce additionally holds ISO 27001, 27017, and 27018 certifications. Agentforce is also HIPAA-eligible under Salesforce's Business Associate Addendum.
  • Data retention: Agentforce runs on Salesforce's Einstein Trust Layer, which maintains zero-data-retention agreements with third-party LLM providers used behind the scenes — your data isn't retained or used for training by those providers when accessed through Agentforce.
  • Unique measures: the Trust Layer adds dynamic grounding (pulling only the specific data a request needs, rather than broad access), data masking, prompt injection defense, toxicity scoring, and a full audit trail of every AI interaction — all inside your Salesforce org's existing trust boundary.

What ToolboxAI does on top of this​

  • Agentforce calls never leave your org. ToolboxAI calls your agent through Salesforce's in-org invocable action rather than an external HTTP callout, so there's no separate OAuth app, Named Credential, or external network hop for Agentforce traffic at all.
  • API keys are stored as encrypted credentials, not in a custom field or custom setting — OpenAI and Claude keys are saved into Salesforce Named Credential / External Credential principals, which Salesforce encrypts at rest and never exposes back through the UI or API once saved.
  • PII masking before anything is sent to a provider. When data masking is enabled (see ToolboxAI Data Options), ToolboxAI scans record fields, activities, and related records for PII (like emails and phone numbers) and pseudonymizes or redacts it before the prompt is ever sent to OpenAI or Claude.
  • Full auditability inside Salesforce. Every AI call — the prompt sent, the response received, the model, token counts, and the record it relates to — is logged to a record your org controls, so you can review exactly what was sent and returned without depending on the provider's own logs.