このページは英語版のみご利用いただけます。 英語版が正式な内容となります。
Security model
What MZGrid's container isolation and contracts protect, what they do not, how to handle keys and secrets, on-chain risks and how to report a vulnerability.
This page describes what MZGrid protects, what it does not, and what you should do to keep your workloads, keys and funds safe. Read it before running anything sensitive on the network.
The most important point
警告
MZGrid jobs run on third-party hardware operated by independent providers. MZGrid provides isolated execution, not confidential computing. A provider with physical or root access to their machine can, in principle, inspect the memory, disk and network traffic of containers running on it. Do not send data, model weights or credentials to a job unless you are comfortable with that possibility.
Confidential-computing technologies (hardware-encrypted memory and remote attestation) are not part of MZGrid's guarantees today. Choose your workloads accordingly: public or non-sensitive models and data are a good fit; regulated personal data, trade secrets and long-lived production credentials are not.
Container isolation
The provider agent runs each job in its own Docker container with the following restrictions:
| Control | Effect |
|---|---|
--gpus device=<uuid> | The container sees only the GPU assigned to the job. |
| No host network | Containers use an isolated network; only the template's declared ports are published. |
| Read-only root filesystem where possible | The image cannot be modified at runtime; writes go to scratch space. |
| Per-job tmpfs | Scratch data lives in memory-backed storage that disappears when the job ends. |
| Resource limits | CPU, memory and process limits are applied per container. |
| No privileged mode | No --privileged and no added capabilities. |
| Secrets via environment only | config.env values are passed only to that job's container and are not written to agent logs. |
--rm | The container is removed when the job ends. |
What isolation protects
- Customers from each other. Jobs on the same machine cannot see each other's processes, filesystems or GPUs.
- Providers from customers. A customer workload cannot reach the host network, escalate privileges through the container runtime configuration, or persist on the host after the job.
What isolation does not protect
- Customers from the provider. The machine operator controls the kernel, the driver, Docker and the hardware.
- Data in transit to the endpoint. Endpoints are served directly from the provider's machine. Unless your workload terminates TLS itself, traffic to
endpointUrlis not encrypted end to end, and anyone who learns the URL can send it requests unless you configure authentication (for vLLM templates, setVLLM_API_KEYinconfig.env). - Correctness of every individual result. Benchmarks, usage reports and 2% canary spot checks make cheating detectable and expensive, but they are statistical deterrents rather than proofs for every job.
Protecting secrets in jobs
- Prefer workloads that need no secrets at all: public images, public models, public data.
- When a secret is necessary (for example a Hugging Face token for a gated model), use a dedicated, read-only, short-lived credential, and revoke it after the job.
- Never pass wallet private keys, cloud root credentials or production database passwords to a job.
- Write results to storage you control using credentials scoped to a single bucket or prefix, and rotate them regularly.
- Prefer providers with a long track record and high reliability for anything you care about; filter with
minReliability.
Account and key handling
Customers
- Funding always requires a wallet signature. Neither MZGrid nor an API key can move tokens out of your wallet.
- Sign permits only for the escrow address on the contracts page and only for the deposit amount shown in the deploy dialog. Check the domain name (
MZGrid) and chain id (8453 on Base mainnet). - API keys (
mzg_live_…) can create and stop jobs and read logs. Keep them server-side, one per service, and revoke any key you suspect is exposed viaDELETE /v1/keys/:id. - Session tokens from wallet sign-in expire; treat them like passwords until they do.
Providers
- The provider key is stored in an encrypted keystore (
~/.mzgrid/keystore.json). Back up the keystore and passphrase separately, restrict file permissions, and keep the passphrase out of shell history. - Keep only the ETH needed for gas on the provider key, and move liquid earnings to a separate wallet you control.
- Anyone with the provider key can request exit and, after 10 days, withdraw your lock. If you suspect compromise, act immediately.
- Keep the NVIDIA driver, Docker, the NVIDIA Container Toolkit and the host OS patched: container isolation is only as strong as the host.
On-chain risk
The contracts are open source and verified on BaseScan; source is published at github.com/mzgrid. Before relying on them, review the verified source yourself. No third-party security audit is claimed on this page.
Risks to be aware of:
- Smart-contract bugs. Despite testing, contracts may contain defects. Because they are immutable, a defect cannot be patched in place.
- Coordinator trust. The coordinator's oracle key reports
usedSecondsand can fail jobs; its slasher key chooses slash amounts up to a provider's locked balance. The contracts bound what these keys can do — settlement can never exceed a job's deposit and only pays the client, provider, registry and treasury — but they do not verify the reported facts. If the coordinator is unavailable, customers can stillcancelJobandreclaimTimeoutdirectly. - Admin multisig. The multisig can grant and revoke roles on the escrow and registry, including the oracle and slasher roles. It cannot change fees, lock rates, delays or caps, which are constants.
- Base network. MZGrid inherits the security and liveness properties of Base. See docs.base.org.
- Token risk. MZG's market value may fluctuate. See the risk disclosure.
Responsible disclosure
If you believe you have found a vulnerability in the MZGrid contracts, coordinator API, provider agent, SDK or website, email [email protected] with:
- a description of the issue and its impact,
- steps or a proof of concept to reproduce it,
- affected components, contract addresses or versions,
- how you would like to be credited.
We follow a 90-day coordinated disclosure policy: please give us 90 days from your report — or until a fix is released, if sooner — before disclosing publicly, and we will keep you updated on progress during that period. Do not access or modify other users' data, disrupt the service, run tests against other users' jobs, or exploit an issue beyond what is needed to demonstrate it. Do not use public channels such as Discord or GitHub issues for security reports.
Operational status
Service health for the API, scheduler, chain indexer and regions is published on the status page.