Security

A stolen token authorizes nothing.

Oathera is built so that what an attacker can copy is not enough. The agent's key stays on its host, tokens are short, every call carries a proof, and the gateway refuses what it can't verify.

A stolen token is refused without the session key An attacker copies a token from a log and sends it to the gateway. The gateway checks for a single-use proof signed by the session key, which the attacker does not hold, so the call is refused. a log token attacker replays it Agent Enforcement needs a proof REFUSED no session key, no proof A STOLEN TOKEN AUTHORIZES NOTHING
The security model

Safe when something leaks.

Keys stay on the host

The anchor key is created on the agent's host, in a file only the agent's user can read. Each session key lives only in memory and is never written to disk.

Short tokens

Identity tokens last five minutes at most, and each token request is signed by the anchor key and used once.

Proof of possession

Every call carries a single-use proof signed with the session key the token names. A token copied out of a log authorizes nothing, and a replayed proof is refused.

Fail closed

Oathera Agent Enforcement refuses a call when the token, the proof or the boundary decision doesn't verify, or when no decision comes back in time.

Host binding

The host is compared with the enrollment record at every token. Enforce refuses a mismatch; warn, the default for testing and design partners, flags it as a high-severity alert.

A guard that checks itself

Both MCPs load their signed protected core only after checking its signature, check it again while running, and stop if it has changed.

Honest by design

Oathera does not verify people.

Your identity provider and your proofing sources say who a person is. Oathera proves things about the agent: which agent is acting, which key signed, which host it runs on, and which principal approved it.

What we don't do

Oathera doesn't judge model output, inspect or rewrite prompts, replace your identity provider, or verify people itself.

Sponsor and clearance

A named, verified sponsor for each agent, and clearance above the baseline of 0 for an agent whose evidence supports it.

Evidence grades

Evidence is graded as self-reported, corroborated or verified, so self-reported evidence never passes for verified. Host attestation status and host confidence are recorded at enrollment.

Revocation

Stop one agent without touching anything else.

Revoke an agent

Its next token request and its next boundary decision are refused. A token already issued stays valid until it expires, five minutes at most.

Deprovision a person

Stop every agent a principal approved, without touching the person's own sign-in.

Revocation snapshots

Signed revocation status, refreshed faster for agents with a more dangerous risk tier.

Audit

Every security decision, linked in order.

Hash-chained decisions

The Oathera Identity Service links each security decision into a hash chain. The chain is held in memory, and the hosted demo resets it when its container sleeps.

No secrets in the log

Credential-like values are replaced by a correlation ID before they are logged.

Sealed and exported

Signed batch seals, chain verification, and export exactly once to your SIEM or a write-once store.

Keys

Key custody and rotation.

Anchor key on the host

A software key in a file only the agent's user can read, created by Oathera Agent Enrollment.

Hardware or KMS custody

The anchor key held in a key manager, hardware module, TPM or secure enclave.

Anchor rotation

Move an agent to a new anchor key without enrolling it again.

Service signing keys

The Oathera Identity Service signs with one active key at a time. Rotation keeps the old key in the published key set while the new one takes over.

Operator access

Who runs the console.

Console sign-in and team roles

The Oathera console has its own sign-in, admin and viewer roles, and a step-up code to create a team.

Oathera Identity Service admin roles

Role checks with separation of duties, a fresh sign-in for high-risk actions, and agent tokens refused on the admin interface.

Ask us the hard questions.

Talk to the founders about your threat model, or read the developer section for the detail behind every claim on this page.