Skip to main content

Zero trust

Never assume trust. Every request is authenticated and authorised in real time, against your policies, before anything executes.

Why agents make this urgent​

Applying zero trust to people is well understood. Applying it to agents matters more, for reasons specific to how agents behave:

  • Volume. An agent may act thousands of times an hour. A permission granted once is exercised continuously.
  • No suspicion. A person receiving an odd instruction hesitates. An agent does not. It will execute a compromised instruction with the same diligence as a legitimate one.
  • Speed. By the time a human notices, an agent has completed the work.

So the question is not "is this agent trusted?" but "should this action, on this system, be permitted right now?"

Continuous verification​

Unlike traditional automation, which authenticates once at the start of a run, Automentic agents re-authenticate at critical steps.

A workflow that runs for an hour and authenticates once has a one-hour window in which a compromised session stays valid. Re-authenticating before each consequential step shrinks that window to the gap between steps.

Practically:

  1. The agent presents its identity. If it cannot, nothing proceeds.
  2. The request is evaluated against policy — this action, this target, now.
  3. Only then does the action execute, and get signed.
  4. Before the next consequential step, the check happens again.

Revocation that takes effect​

A zero-trust claim is only as good as how fast you can withdraw trust. Because identity is checked at each critical step rather than cached for the run, revoking an agent stops it mid-workflow — not at the next restart.

That is the property to test during an evaluation: revoke an agent while it is working and watch what happens.

The human stays accountable​

Autonomy is not the goal; leverage is. Every consequential action traces back to a person who authorised the policy that permitted it. The agent is the instrument, not the excuse.

For human oversight, Automentic supports FIDO2 and WebAuthn so approvals are phishing-resistant.