Skip to content
AgentIDDocs

Type at least two characters to search titles, headings and content.

↑ ↓ to navigate↵ to open
AgentID Docs

Data protection

What is logged and encrypted

What AgentID stores for each event, what the AI provider and the logs receive, and how encrypted storage and Reveal work.

Per event

DataStored
Prompt and answerThe text as checked and logged: masked when masking is on (browser extension prompts are stored redacted in any case). With Store Encrypted Payloads on, the views show a redacted value and the content is stored encrypted
Originals behind placeholdersOnly with Store Encrypted Payloads on, encrypted, and only visible through an audited Reveal
What the provider received per masked valueFor browser extension turns, the token and the value sent (the token itself)
Detection resultsDetected entity types, security signals, the guardrail decision and its reason
Identityuser_id, the API key's name, prefix, owner label and environment, and request identity hints you send
Model and usageModel id, token counts, latency and cost (from your usage data)
AttachmentsFile names, types, extraction results and the masked text; never the original file after protection

What leaves your side

RecipientReceives
AI providerThe protected text with placeholders; never the placeholder mapping
AgentID logsThe masked text, plus encrypted originals only if you enable encrypted storage
The userThe answer with originals restored on their device, if restoration is on

Guard itself must read the original text to detect sensitive values. It processes the text in memory for the check; what it keeps afterwards is listed above.

Encrypted storage

When Store Encrypted Payloads is enabled in Guardrails:

  • prompts, outputs and the token-to-original mapping are encrypted before they are stored, with AES-256-GCM and a separate data key per event,
  • list and detail views show redacted placeholders instead of content,
  • if encryption is required but not available, AgentID refuses to store the event instead of storing plaintext.

Reveal

Admins of an AI system can Reveal payload on an encrypted event in the activity log. Each Reveal is written to the audit log as ENCRYPTED_EVENT_PAYLOAD_VIEWED, recording whether the input, the output and the masking originals were revealed. Events erased under a GDPR request cannot be revealed.

Local data on user devices

IntegrationStored locally
Browser extensionPlaceholder mappings per conversation, encrypted with a non-extractable key in the extension's own storage, for 30 days after last use; provider pages cannot read them
Claude pluginAn encrypted per-session vault (AES-256-GCM) for 7 days; the device credential in the OS credential store (DPAPI on Windows, Keychain on macOS) or an owner-only file
Claude Desktop companionThe device credential (DPAPI on Windows, Keychain on macOS)

Webhooks

Security events sent to your webhooks carry SHA-256 hashes of the input and output instead of the text, plus the event metadata. For systems that block personal information, the payload is redacted before delivery.

See also Security and privacy.