Skip to content
AgentIDDocs

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

↑ ↓ to navigate↵ to open
AgentID Docs

Security & privacy

Security and privacy

How AgentID protects credentials, data in transit and at rest, and the data that passes through it.

Data minimisation by design

  • With masking on, AI providers receive typed placeholders instead of personal data and secrets.
  • Restoration happens only on the user's device or in your own application. The AI provider and AgentID's logs keep the placeholders.
  • Placeholder mappings in SDK wrappers live in memory for the request. The browser extension and the Claude plugin keep them encrypted on the device.
  • Protected files are replaced with masked text copies; the original file is not uploaded to the AI after protection.

Credentials

  • API keys are stored as hashes; only the prefix is visible after creation. Keys are scoped to one AI system and can be revoked individually or all at once.
  • Browser rollouts use time-limited enrollment tokens (default 7 days, at most 30) that are exchanged for a dedicated key, so users never handle a long-lived key.
  • Claude installations each get their own revocable credential. Plugin bundles contain no API key, only a revocable activation.
  • On devices, credentials are kept in the OS credential store where available (Windows DPAPI, macOS Keychain) or in owner-only files.

Transport and storage

  • All API traffic uses HTTPS.
  • Stored data is encrypted at rest by the hosting platform.
  • With Store Encrypted Payloads, prompts, outputs and the originals behind placeholders are additionally encrypted by AgentID before storage (AES-256-GCM with a separate data key per event). If encryption is required but unavailable, AgentID refuses to store the event rather than storing plaintext.
  • Staged document uploads are private, one-time, and deleted after their text is extracted.

Access control and audit

  • Organization and system roles limit who can change guardrails, manage keys and view sensitive areas; see Organization and access.
  • Decrypting stored content (Reveal payload) is limited to system admins and is recorded in the audit log.
  • Webhook deliveries are signed with HMAC-SHA256 and carry hashes instead of prompt text.

Your responsibilities

AgentID protects the traffic that passes through it. To get full coverage:

  • route every LLM call through an AgentID wrapper or an explicit Guard → provider → Ingest flow,
  • protect the full message history you send to a provider, not only the latest message,
  • keep API keys server-side and rotate them when people or services change,
  • choose fail-closed for systems where unchecked data must never reach a provider,
  • use content policies for sensitive terms only you know.

Compliance

AgentID provides building blocks for GDPR and EU AI Act programmes: data minimisation through masking, encrypted storage with audited access, transparency metadata, instructions for use and technical documentation per AI system (see Compliance). Compliance of your use of AI also depends on your own processes, contracts and records.