🛡️ Security

Security principles at Fume are shared across products: restrict access to the minimum each party needs, make every sensitive action traceable, and keep data where it belongs. This page covers the AI agents first, then the Tokenization Engine.

AI agents

Agents act inside limits you set, and every action is logged and auditable, so your compliance team can trace each run. We monitor every run in production and handle the edge cases rather than leaving them to your team.

Agent engagements run under a written information security and data protection policy with a named executive owner, and we share it with clients so they can assess our posture directly. Before any client data flows, we sign a data processing agreement, and each engagement gets its own security overview describing how data moves through the solution, where it is processed, and how long it is kept.

We keep the attack surface small. Client data processing runs on a short list of established infrastructure providers holding ISO 27001 or SOC 2 certification, with EU-first hosting, encryption in transit and at rest, least-privilege access, and multi-factor authentication on every account that can reach client data. Where a language model reads a document, the call goes through an AI gateway with zero data retention enforced: processing is transient, and client data is never used to train models, ours or anyone else's.

Retention is bounded, and at the end of an engagement client data is deleted or returned, at the client's choice. Agents work from written, version-controlled instructions derived from the client's own documented process. Their output goes to a human for review, and an agent that hits something its instructions do not cover stops and escalates rather than improvising. Automations also start out running alongside the client's existing manual process, so a documented fallback is always in place.

For funds whose data can't touch a public AI cloud, we deploy the same agents on open-source models entirely inside your own infrastructure. See Private & Local AI.

Tokenization Engine

Everything from here down is about the Tokenization Engine and the platform around it.

The protocol is designed so that underlying liquidity stays in secure, segregated custody, and only interacts with the protocol at key moments: subscriptions, redemptions, and fee management. After a fund is deployed, only authorized parties can interact with its smart contract: the fund manager, and investors the manager has whitelisted. This drastically reduces the on-chain attack surface.

We engaged ChainSecurity, one of the leading smart contract auditors, to review the protocol. Read the report.

For the full picture (implementation practices, testing, upgradability, and the catch-up mechanism), see Security & Risk Mitigation in the whitepaper.

Authentication and data handling

There are no passwords on the Tokenization Engine. Users sign in with a one-time code sent to their email, powered by Privy (security details). Logging in gives a read-only view; any action that moves money or changes a fund requires wallet approval from the fund manager, and we recommend multi-signature solutions like Safe Wallet or Fireblocks for the governance wallet.

We store nothing on-chain that could be considered personal or sensitive data. That data lives on our servers, protected by industry-standard security and encryption. All wallets with important responsibilities on the Fume side are multisig smart wallets.