Skip to content

security.mdx

Latest commit

 

History

History
74 lines (54 loc) · 3.94 KB

security.mdx

File metadata and controls

74 lines (54 loc) · 3.94 KB
Raw
title Security
description AI agent security and governance. An agent cannot act beyond what you allowed, and a proof never exposes the data behind it.
icon lock

Proofable checks what is required, shares only what a decision needs, and records the decision instead of the sensitive data behind it.

Invariants

These hold everywhere in Proofable. A deviation from any of them is a bug.

  • No implicit authority. Every action is checked against identity, proof, policy, and the permission an owner granted, at the moment it runs. Identity, ownership, sign-in, or possession of a key never grants authority by itself.
  • Minimum disclosure. A check returns the minimum claim or result the decision requires.
  • Private and offchain by default. Public exposure, indexing, IPFS, and chain anchoring are explicit choices.
  • Secrets never become agent context. Connected tools run without handing credentials to the model. Credentials are resolved inside trusted server execution and never placed in a tool schema, proof, browser response, or public record.
  • Evidence stays private. Private evidence can exist to keep a proof auditable. Public readers get the claim, status, or proof, never the source material.
  • Every privileged action has a policy decision. Subject, resource, action, policy, and context produce allow, deny, or approval. The proof records the decision.
  • Permission is narrower than ownership. Agents act only inside the permission they were given. Spend, secrets, tool access, and high-risk actions stay bounded.
  • Auditability without surveillance. Records show who authorized what, under which policy, against which proof, and what happened. They are not a store for request bodies, prompts, OAuth payloads, or provider responses.

One model for every decision

Sign-in, connected apps, payments, agents, and jobs all follow the same path:

Evidence → Proof → Policy → Authority → Action → Receipt
  1. Sensitive evidence is examined inside a trusted boundary.
  2. It is reduced to a claim or proof.
  3. The claim is checked against policy and the permission an owner granted.
  4. The action is allowed, denied, or sent for approval.
  5. A proof records what was decided.
  6. The evidence stays protected, or is deleted according to its retention rule.

Some features need private state, encrypted secrets, or private evidence. Proofable keeps that data purpose-bound, least-privilege, auditable, and deletable where applicable.

Proof visibility

Proofs are private unless you choose otherwise. Visibility modes and defaults are in Proofs.

Smart contract audit

SafeStack AI audited the Proofable smart contracts in March 2026. The audit found no critical, high, medium, or low severity issues.

The scope covered the verifier registry, voucher hub, voucher spoke, and token contracts. Status is listed on the Trust Center. Email dev@proofable.me for the full report.

Integration checklist

  • Decide access with a server-side gate check, not a browser flag.
  • Respect each check's freshness window.
  • Run interactive checks through Hosted Verify at /verify.
  • Read the verifier list from the live catalog.
  • Keep proofs private unless your product needs them public.
  • Keep access keys and other secrets on servers.

Report a vulnerability

Email dev@proofable.me and do not post it publicly. The full policy is in Security disclosure.

Policies

What Proofable processes and why. The terms for using Proofable. Copyright notices and takedowns.