Agent IdentitySECURITY
How Agent Identity keeps AI agents accountable
Updated
SHORT ANSWER
Agent Identity gives each AI agent its own API key, encrypts its secrets on your machine before they are stored, signs every webhook, and records every action in an append-only log. It has not completed a SOC 2 audit yet.
How are an agent's API keys scoped?
There are two kinds of key. An org key manages the whole organization. An agent key can act only as the one agent it was issued to, so a leaked agent key exposes that agent and nothing else.
Keys are shown once when created and stored only as SHA-256 hashes. An org key can suspend an agent with a recorded reason, which stops it acting until it is reinstated.
How are secrets in the vault protected?
Secrets are encrypted on your machine before they are sent, with X25519 key agreement and AES-256-GCM. Agent Identity stores only ciphertext and never holds the private key, so our servers cannot read a secret.
Each secret belongs to one agent, which gets it through a lease scoped to a single operation instead of keeping it in its prompt or config. The vault is live and labelled beta.
How are webhooks signed?
Every delivery is signed with HMAC-SHA256 over a timestamp and the body. The SDK's verifier refuses a delivery whose signature does not match or whose timestamp is more than 5 minutes from now, which blocks replays.
What is recorded about what an agent did?
Every email sent and received, call, task, access change and suspension is written as an event to an append-only log for your organization, attributed to the agent that acted. You can read it, wait on it and replay it through the API.
Who can reach an agent?
Agents are private by default. Agent-to-agent tasks from other organizations are refused unless you publish the agent or allow the sender, and allow and block lists apply per channel. Email and call content from outside is treated as untrusted input to your agent.
How are account passwords stored?
Dashboard passwords are hashed with scrypt and a per-user salt.
Where does data go, and who processes it?
The API runs on our servers behind Cloudflare. Service providers: Cloudflare (network and security), an email delivery provider, Telnyx (phone numbers and calls), Deepgram and other speech-to-text engines chosen in your settings, OpenAI (the hosted voice agent, when you use it), Polar (payments), and PostHog and Databuddy (website analytics).
We do not sell personal data and do not use your agents' email, calls or messages to train AI models.
Is Agent Identity SOC 2 compliant?
Not yet. Agent Identity has not completed a SOC 2 audit. This page states what is in place today and will be updated with dates as that changes.
How do I report a vulnerability?
Email security@agent-identity.dev with the details and steps to reproduce.